Abstract:With the rapid development of 3D vision, large-scale 3D point cloud processing in real time based on deep learning has become a research hotspot. Taking a large-scale 3D point cloud with disordered spatial distribution as the background, this study comprehensively analyzes, introduces and compares the latest progress of deep learning in real-time processing of 3D vision problems. Then, it analyzes in detail and compares the advantages and disadvantages of algorithms in terms of point cloud segmentation, shape classification and target detection. Further, it briefly introduces the common data sets of point clouds and compares the algorithm performance of different data sets. Finally, the study points out the future research direction of 3D point cloud processing based on deep learning.

Abstract:Blockchain technology has brought new changes to cryptocurrencies and has been widely used. However, it still faces the needs and goals of high throughput, low transaction latency, security, and decentralization. In addition, the willingness of consumption nodes (i.e., transaction providers) is difficult to be mapped into leaders, and block miners are keen on mining competitions, which also leads to intensified centralization and energy consumption. To this end, a new consensus algorithm, PoM (proof-of-market), and its first implementation case, the Achain protocol, are proposed. The algorithm is different from the traditional PoW (proof-of-work) consensus, and its design enables consumer nodes to perform PoW and vote for leader nodes, which not only discretizes the mining, improves decentralization, and reduces energy consumption but also reflects the willingness of consumer nodes. In other words, only the node mainly supported can become the leader. In terms of performance, Achain also improves scalability compared with PoW-type blockchains, and it provides a solution for node storage and optimization, which is called FastAchain. In terms of security, Achain is supplemented by a set of incentive-compatible reward and punishment mechanisms to make malicious nodes have negative revenue expectations, which protects the interests of honest nodes, and Achain can tolerate up to 1/3 of the total network computing power being controlled by the malicious nodes. In order to verify Achain’s performance, a prototype of Achain under a large-scale network is implemented for evaluation. The results show that Achain has achieved expectations, outperformed some mainstream representative blockchain protocols, and maintained positive chain convergence and decentralization.

Abstract:Accurate prediction of the future trend of the commercial sales volume is of great importance to the development and operation of enterprises and the macro-control by the government. Traditional data prediction methods are time-consuming and subjective, while the existing data-driven future business prediction methods do not take into account the diversity of features in the data sets. The data of the commercial sales volume is time-series, which contains a wealth of time window features, lagging historical features, and price change trend features. Previous studies tend to focus only on some of these features, and the integration and enhancement of these features are seldom explored. The prediction accuracy of the existing future business prediction methods still needs to be improved. Therefore, this study proposes a future business forecast method based on multimodal feature aggregation, which firstly preprocesses the commercial sales volume data and then extracts five different groups of time window features and other features of the data set on the basis of feature engineering. In machine learning, the hard voting mechanism is used to select the appropriate model for the training of the five groups of time window features. At the same time, the neural network optimization model is applied to extract the time-series features and forecast results, and then, the dependency relationships between the data set of the sales volume and some features are analyzed. Finally, with the soft voting model, a high-precision forecast of the commercial sales volume is achieved by complete model integration. The experimental results reveal that the proposed method has high prediction accuracy and efficiency, which is greatly better than the existing prediction methods.

Abstract:This study is conducted to measure the level of industrial green development and judge the differences in industrial green innovation ability among provinces. The study constructs a multi-index evaluation system integrating comprehensive attribute values for industrial green development from the perspective of green input factors and green output benefits and proposes an interval multi-attribute measurement method based on the hybrid multi-dimensional cloud model. This method creatively uses the mutual transformation of the interval weight and cloud weight to solve the problem of different weights of multiple indexes. After that, the parent cloud closeness is employed to calculate the industrial green development level, and the cloud projection is applied to measure industrial green innovation ability. Finally, industrial panel data between provinces are used for verification. The results indicate that compared with the results of the general multi-index comprehensive evaluation method, the empirical results of this method are more consistent with the actual situation. This means that the method can not only evaluate and analyze the overall situation of the industrial green development level but also accurately calculate the contribution of each index to judge whether a province or region has the industrial green innovation ability. Therefore, this study can provide substantive suggestions and theoretical decision-making basis for regions to adjust the measurement indexes for the industrial green development level and formulate industrial green development plans.

Abstract:The existing research on battery state of charge (SOC) prediction based on neural networks mostly focuses on the optimization of model structure and related parameters, ignoring the important role of training data. A battery SOC prediction method based on feature selection and data augmentation is proposed to overcome this problem. Specifically, feature engineering is carried out according to the original battery charge and discharge data, and seven features that are most helpful to model prediction are selected by the permutation importance (PI) method; then, Gaussian noise is added to expand the total number of training data samples and thereby achieve the purpose of data augmentation. In the experiment, a bidirectional long short-term memory (Bi-LSTM) network is used as the prediction model, and the Panasonic 18650PF dataset is adopted as the training data. When the standard Bi-LSTM model is employed for prediction, the mean absolute error (MAE) and the maximum error (MaxE) are 0.65% and 3.92% respectively. After feature selection and data augmentation, the MAE and MaxE of model prediction are 0.47% and 2.62% respectively, indicating that the accuracy of the battery SOC prediction model can be further improved by PI feature engineering and the Gaussian data augmentation method.

Abstract:Facial expression recognition is easy to lose a lot of useful feature information during feature extraction and cannot extract more comprehensive facial expression features. In view of these problems, a multi-scale feature fusion network model (DS-EfficientNet) is proposed. The model includes a deep network and a shallow network. The shallow network is used to extract the detailed texture information of facial expressions, and the deep network is used to extract the global information of expressions. An attention mechanism is added to the shallow network to enhance the ability to extract shallow detail information. Finally, feature fusion is performed on channels, and the network can extract more abundant facial expression information after the fusion. In order to reduce the model parameters and improve the generalization performance of the model, the fully connected layer is replaced by a global average pooling layer, and batch normalization is added. The method proposed in this study is tested on Fer2013 and CK+, and the recognition accuracy reaches 73.47% and 98.84%. Experiments show that this method can extract more abundant facial expression information, and the model has a strong generalization ability.

Abstract:In the all-media era, recommendation based on multimodal data is of great significance. This study proposes recommendation based on data in three modalities: text, audio, and image. Tensor fusion is implemented in two stages: The correlation between any two modes is modeled and fused by three parallel branches in the former stage, and the results of the three branches are then fused in the latter stage. This approach not only considers the local interaction between two modalities but also eliminates the influence of the modality fusion order on the result. In the recommen-dation module, the fused features are input to the stacked denoising auto-encoder and are then used as auxiliary features of collaborative filtering for recommendation. In the recommendation system constructed, an end-to-end training process is adopted for modality fusion and recommendation. Moreover, to overcome the high similarity and poor diversity of the recommendation results, this study also constructs a similarity matrix with the fused features of the tensor modalities in the two stages to further refine the results on the basis of the available recommendation results and thereby achieve rapid diversified recommendation. The experimental results show that the recommendation model based on the proposed multimodal fused features can not only effectively improve recommendation performance but also enhance the diversity of recommendation results.

Abstract:At present, the good performance of deep learning in medical image analysis mostly depends on high-quality labeled datasets. However, due to the professionalism and complexity of medical images, the labeling of datasets often requires huge costs. To tackle this problem, this study designs a semi-automatic labeling system based on deep active learning. This system reduces the number of labeled samples required for the training of the labeling model based on deep learning through the active learning algorithm, and the trained labeling model can be used for labeling the remaining dataset. The system is built on the basis of a Web application, which does not require installation and can be accessed across platforms. It is convenient for users to complete the labeling work.

Abstract:In response to different noises in iris images, such as occlusion by glasses, blur, and angle deviation, this study designs a convolutional neural network (CNN) embedded with Mask R-CNN, named Mask-INet, for iris segmentation. The network adds a bottom-up path to the feature pyramid in the feature extraction stage, which not only improves the localization information of bottom-to-top features and enhances semantic information fusion but also further accelerates bottom-to-top propagation efficiency and effectively improves the accuracy of iris feature extraction. To further explore the feature information in the feature map, the study introduces upsampling and a convolutional block attention module (CBAM) network in the mask prediction branching stage. Upsampling is used to improve the spatial resolution of the feature map, and the CBAM network helps make the salient information in the feature map more significant so as to enhance the discrimination capacity for the features. The method is validated on the iris dataset provided by the NIR-ISL 2021 competition. The method outperforms the network of the champion of the event in terms of all indicators under the same experimental conditions. Compared with the baseline Mask R-CNN, the proposed method has the Dice similarity coefficient, mean intersection over union (mIoU), and recall improved by 8.53%, 11.97%, and 8.88%, respectively, which boosts iris segmentation performance.

Abstract:In order to improve the recognition rate of plant leaf disease detection by edge computing devices, this study uses a convolutional neural network to build a plant leaf target recognition model and a plant leaf disease classification model and adopts OpenCV to integrate the two models into a plant leaf disease detection system. The target areas of plant leaves are positioned and clipped by the single shot multibox detector (SSD) algorithm, and then the plant leaf disease classification model is used to classify the clipped plant leaf areas according to diseases. At the same time, the classification model is optimized by TensorRT accelerated inference. In addition, on the same host device and Jetson Nano computing platform, a comparative experiment is carried out on the model before and after optimization. The experiment shows that the recognition rate of the optimized plant classification model on the same host device increases by 22 times. At the same time, the optimized classification model improves the recognition rate of the plant leaf disease detection system by seven times. Furthermore, the optimized system is deployed on the Jetson Nano computing platform, and the detection rate of plant leaf diseases is increased by 10 times compared with that before optimization, which thus realizes real-time plant leaf disease detection.

Abstract:As manual identification of ground-penetrating radar (GPR) pipeline images faces the problems of low efficiency, large errors, and high costs, this study proposes an intelligent pipeline target detection method based on improved Cascade R-CNN. First, the GPR pipeline image data set is preprocessed to improve data quality. ResNeXt is used instead of ResNet as the backbone network to extract target feature information, and a multi-scale feature fusion module FPN is added to fuse high-level features to low-level features to enhance the expressiveness of low-level features. Secondly, the Gaussian non-maximum suppression (NMS) method Soft-NMS is used to obtain more accurate candidate boxes, and Smooth_L1 is taken as the loss function, which accelerates model convergence and reduces the probability of gradient explosion during training. Finally, for the special shape features of the pipeline target, the appropriate aspect ratio and size of the anchor boxes are set to improve the quality of generated anchor boxes. The experimental results demonstrate that the proposed method achieves the intelligent detection of underground pipeline targets with complex features, and the average accuracy of target detection reaches 94.7%, which is 10.1% higher than that of the Cascade R-CNN method.

Abstract:The single image super-resolution (SISR) reconstruction algorithm is ill-posed in the mapping learning from low-resolution (LR) image to high-resolution (HR) image, and the deep neural network has slow convergence and lacks the ability to learn high-frequency information. Moreover, image feature information tends to be lost during deep neural network propagation. In order to address these issues, this study proposes an image super-resolution reconstruction network based on dual regression and residual attention mechanism. Firstly, the mapping space is constrained by dual regression. Secondly, a residual attention module (RCSAB) is constructed by combining channel and spatial attention mechanisms, which not only accelerates the model convergence speed and effectively strengthens the learning of high-frequency information. Finally, a dense feature fusion module is introduced to enhance the fluidity of feature information. In addition, a comparison with the mainstream SISR algorithms is carried out on four benchmark datasets, namely, Set5, Set14, BSD100, and Urban100, and experimental results demonstrate that the proposed method is superior to other algorithms in terms of objective evaluation metrics and subjective visual effects.

Abstract:The traditional architecture based on a central server is an important solution for the construction of background services in the past. However, with the explosive growth of the number of users and application requirements, this architecture has put forward higher requirements for the computing and storage capabilities of the central node and brought a crisis of confidence. A typical representative of distributed systems, namely, blockchain, is the core technology of Bitcoin, and it has been widely concerned by researchers in recent years due to its characteristics such as tampering prohibition, traceability, and forgery data prohibition. This study aims to apply the consortium blockchain to the charging network composed of electric vehicles, charging piles, smart electric meters, and transmission networks and use blockchain technology to manage charging records, so as to protect the interests of each party and provide transaction disputes with data-level support. In addition to a dedicated consortium blockchain, this study also proposes a new consensus mechanism and a corresponding smart query contract suitable for the trusted charging model of electric vehicles. The experimental results show that the designed consensus mechanism can operate safely and efficiently in this trusted charging network model and can make users quickly search for transactions.

Abstract:Maintaining a safe social distance is one of the important means to effectively prevent the spread of the virus. Moreover, it can not only reduce the number of infected people and ease the medical burden but also greatly lower the mortality rate. On the basis of the you only look once version 4 (YOLOv4) framework, the lightweight network E-GhostNet is used to replace the CSPDarknet-53 in the original network. The E-GhostNet network establishes a relationship between the input data and the output features generated by the original Ghost module, thereby enabling the network to capture contextual features. Then, the coordinate attention (CA) mechanism is introduced to E-GhostNet to enhance the model’s attention on effective features. In addition, the complete intersection over union (CIoU) loss function is replaced by the soft intersection over union (SIoU) loss function to obtain a faster convergence speed and optimization effect. Finally, the DeepSORT multi-target tracking algorithm is utilized to detect and label pedestrians, and affine transformation (IPM) is employed to determine the violation of the required distance between pedestrians. The experimental results show that the network achieves real-time pedestrian distance detection with a detection speed of 40 FPS and an accuracy of 85.71%, which is 2.57% higher than that of the original GhostNet algorithm.

Abstract:YOLO is one of the most important algorithm models in the target detection of computer vision. Given the existing YOLOv5s model, an extended YOLOv5 algorithm model for multi-level classification target detection is proposed. Firstly, the function of the annotation tool LabelImg is extended to construct multi-level classification label files. Secondly, the output format of the detection head is modified on this basis of the YOLOv5s algorithm, and the core design idea of the DenseBlock and Res2Net network model is introduced in the front end of the backbone network to extract rich mul-ti-dimensional feature information, enhance the reusability of feature information, and realize the task of YOLO-based multi-level classification target detection. The helmet color is taken as the secondary classification for training and verification on the open source helmet data set, and the average precision, precision, and recall reach 95.81%, 94.90%, and 92.54%, respectively. The experimental results verify the feasibility of the YOLOv5-based multi-level classification target detection task, and the proposed model provides a new idea and method for target detection and multi-level classification target detection tasks.

Abstract:The original damage-free pavement image is of great significance for analyzing the evolution details of pavement damages and formulating the next maintenance plan. However, the initial state corresponding to a pavement crack image cannot be obtained in field acquisition. To obtain the corresponding damage-free pavement image, this study proposes a deep image prior-based unsupervised crack image inpainting algorithm for asphalt pavements that enables efficient semantic-level inpainting of cracks in a single pavement image. Specifically, a robust principal component analysis algorithm is used to remove the vertical stripe noise on the surface of the pavement crack image. Then, the maximum between-class variance method and morphological processing are employed to obtain a binary mask image of the crack area. Finally, the crack area is inpainted with the proposed deep image prior-based inpainting algorithm to obtain the final damage-free pavement image. The proposed method is evaluated on a dataset of self-collected pavement crack images. The experimental results show that the proposed method can effectively achieve semantic-level inpainting of pavement crack images as it significantly improves the peak signal-to-noise ratio and structural similarity to an average of 43.3823 dB and 0.9834, respectively, compared with those of the traditional methods and it also achieves a high speed.

Abstract:The analysis of numerous and intricate data sets is a highly challenging task, in which the technique to detect outliers in data plays a pivotal role. Capturing anomalies by clustering is the most common method among the increasingly popular anomaly detection techniques. This study proposes an anomaly detection algorithm based on second-order proximity (SOPD), which includes clustering and anomaly detection stages. During clustering, the similarity matrix is obtained by second-order proximity. During anomaly detection, the relationships between points in the cluster and the center of the cluster are employed to calculate the distance of all the points in each cluster generated by clustering from the center of the cluster and capture the anomalous state. The density of each data point is also taken into account to exclude the cases of cluster boundaries. The use of second-order proximity enables the locality and globality of the data to be considered simultaneously, which reduces the number of the obtained clusters and increases the accuracy of anomaly detection. Moreover, this study compares this algorithm with some classical anomaly detection algorithms through massive experiments, and the result shows that the SOPD-based algorithm performs well overall.

Abstract:The feature matching algorithm based on deep learning can produce larger scale and higher quality matching than the traditional algorithm based on feature points. This study aims to obtain a wide range of clear pavement crack images and solve the problem of missing matching pairs in weak texture image mosaics. The road image mosaic is realized based on the deep learning LoFTR (detector-free local feature matching with Transformers) algorithm. Given the characteristics of road images, the local mosaic method is proposed to shorten the running time of the algorithm. Firstly, the segmentation of adjacent images is conducted, and the dense feature matching is produced through the LoFTR algorithm. Secondly, the homography matrix value is calculated according to the matching results and the pixel conversion is realized. Thirdly, images after local mosaics are obtained through the image fusion algorithm based on wavelet transform. Finally, some images that are not input into the matching network are added to get the complete mosaic result of adjacent images. The experimental results show that, compared with methods based on SIFT (scale-invariant feature transform), SURF (speeded up robust features), and ORB (oriented FAST and army), the proposed method has a better effect on road image mosaic and higher confidence of matching results in feature matching stage. For the mosaic of two road images, the time consumed by the local splicing method is shortened by 27.53% compared with that before the improvement. The proposed mosaic scheme is efficient and accurate, which can provide overall disease information for road disease monitoring.

Abstract:As the main technical means of computer protection, intrusion detection technology has been widely studied due to its advantages of strong adaptability and ability to identify new types of attacks. However, the recognition rate and false alarm rate are difficult to guarantee, which is the main bottleneck of this technology. To improve the recognition rate and reduce the false alarm rate of anomaly detection technology, this study proposes a terminal-level intrusion detection algorithm (TL-IDA). In the data preprocessing stage, the terminal log is cut into continuous and small-block command sequences, and common statistical indicators are introduced to construct feature vectors for the command sequences. Then TL-IDA is applied to model users through the feature vectors. On this basis, a sliding window discrimination method is also proposed to judge whether the system is under attack, so as to improve the performance of the intrusion detection algorithm. The experimental results show that the average recognition rate and false alarm rate of the TL-IDA are 83% and 15%, respectively, which are superior to those of similar terminal-level intrusion detection algorithms based on anomaly technology such as ADMIT and hidden Markov model.

Abstract:In the process of automated crane operations for port containers, the detection of container truck heads is an indispensable link. To solve the problem of low efficiency by manual confirmation and high costs and complex systems by the laser scanning method, this study proposes an algorithm based on video images of operation scenes and deep learning for target detection of container truck heads. Specifically, upon the construction of a sample data set of container truck heads, the DCTH-YOLOv3 detection model is used, and sample training is performed through the method of model migration learning. The DCTH-YOLOv3 model is an improved YOLOv3 model proposed in this study. The algorithm improves the FPN structure of YOLOv3 and proposes a new feature pyramid structure—AF_FPN. During the fusion of higher- and lower-order features, the AFF module with the attention mechanism is introduced to focus on effective features and suppress interference noise, which increases the accuracy of detection. In addition, the metric CIoU loss is used to replace L2 loss to provide more accurate boundary box change information and further improve the model detection accuracy. The experimental results indicate that the detection rate of DCTH-YOLOv3 can reach 46 fps on GTX1080TI, which is only 3 fps lower than that of YOLOv3. The detection accuracy can reach AP_0.5 0.9974 and AP_0.9 0.4897, in which AP_0.9 is 16.4% higher than that of YOLOv3. Compared with the YOLOv3 algorithm, the proposed algorithm has higher accuracy and can better meet the requirements of automatic operations for high accuracy and fast identification in the anti-collision detection of container trucks.

Abstract:Anomaly detection is one of the research focuses in machine learning and data mining, which is mainly used in fault diagnosis, intrusion detection, and fraud detection. There have been many effective related studies, especially those of the anomaly detection method based on isolation forest, but there are still many difficulties in the processing of high-dimensional data. A new anomaly detection algorithm, k-nearest neighbor based isolation forest (KNIF), is proposed. The method uses hyperspheres as an isolation tool, utilizes the k-nearest neighbor method to construct an isolation forest, and constructs a distance-based outlier calculation method. Sufficient experiments show that the KNIF method can effectively detect anomalies in complex distribution environments and can adapt to application scenarios of different distribution forms.

Abstract:The standard shuffled frog leaping algorithm (SFLA) for optimization has the shortcomings of low optimization accuracy and easy falling into a local convergence area. To improve its performance, this study proposes an improved SFLA (CSFLA) based on local search with a cloud model and cosine leap weight update position. First, Tent chaotic mapping and backward learning are performed to generate a population so that the population has a more uniform distribution. The area where the best individuals in the subpopulation are located is explored by taking advantage of the normal property of the cloud model. Then, the leaping step size weight based on the cosine function is introduced to other individuals in the population, which makes the weight decrease from a high level at different rates throughout the iterations to improve the global search ability of the population. Finally, CSFLA is compared with multiple optimization algorithms on different types of test functions. The results show that CSFLA has a better convergence speed and accuracy and can find the global optimal solution effectively. The improved algorithm is applied to the traveling salesman problem and proved able to find shorter routes.

Abstract:Given low vehicle detection accuracy and serious miss detection in a haze environment, a vehicle detection algorithm with multi-scale feature fusion in a haze environment is proposed. Firstly, the conditional generation and adversarial network is employed to preprocess the haze images. Then, as the object feature is not obvious in a haze environment, a multi-scale feature fusion module is put forward. On the basis of YOLOv3, a shallow branch is added for upsampling splicing and fusing with deep layer features during extracting features from backbone networks. As a result, the feature map with the scale of 104×104 is obtained, which is adopted to enhance the shallow semantic information. The feature enhancement strategy guided by the CBAM attention mechanism is utilized to ensure the integrity of context information and improve detection accuracy. Finally, the dehazed images are sent to the improved YOLOv3 network for detection. Experimental results show that the proposed algorithm has better performance than the YOLOv3 algorithm on the RTTS dataset. The proposed model can achieve an average accuracy of 81% and a recall of 67.52% and can locate vehicles more accurately.

Abstract:Network topology discovery is important for many key network management tasks. However, as the network scale expands, the network structure gets complex. The previous SNMP-based network topology discovery algorithms cannot effectively identify subnet types and multi-IP devices, and they have low topology efficiency and accuracy. In view of the above problems, this study proposes an SNMP-based topology enhancement identification (SNMP-TEI) algorithm. Firstly, the subnet IP address is heuristically determined, and probes are sent to it, so as to judge the subnet type according to the detection results. In addition, probe injection is stopped in a timely manner after the subnet type is determined, so as to prevent the network load from being too large. Secondly, the device fingerprint is set through system information recorded by MIB-II, and a device type identification algorithm is used to identify the device fingerprint of the terminal host IP, so as to identify multi-IP devices. The experimental results show that this method can effectively identify subnets and multi-IP devices in simulated networks and reduce the network load, with a detection accuracy of 96.43%.

Abstract:Video scene switching detection is important in video processing. An excellent algorithm for video scene switching detection is of great significance for subsequent video processing, including information annotation and feature extraction. In this study, on the basis of the traditional image processing methods and oriented FAST and rotated BRIEF (ORB), the similarity estimation of images is realized separately. Through the dynamic fusion factor, the value of similarity obtained by various basic algorithms in the complete algorithm is quantified, and the dynamic threshold is calculated according to the similarity of neighboring frames to improve the stability of the algorithm. In addition, by flow planning, the number of calls of the time-consuming part of the algorithm is reduced, and efficient video scene switching detection is finally realized. The experimental results show that the accuracy of video scene switching detection is improved to a certain extent.

Abstract:With the development of the Internet, how to quickly obtain core information from massive news and make browsing easy has become an urgent problem for information departments. The existing TextRank and its improved algorithm fail to consider text features comprehensively in extracting news summaries. In selecting summaries, they only focus on the redundancy and ignore the diversity and readability of the summaries. In order to solve the above problems, this study proposes a multi-feature automatic text summarization method, namely, MF-TextRank. A more comprehensive text feature information is summarized according to the structure, sentences, and words of news, which is used to improve the weight transfer matrix of the TextRank algorithm and make the sentence weight calculation more accurate. Furthermore, an MMR algorithm is used to update sentence weight, and the candidate summary set is obtained by beam search. According to the MMR score, the candidate summary set with the highest cohesion is selected as the final summary for output. The experimental results show that the MF-TextRank algorithm outperforms the existing improved TextRank algorithm in extracting summaries and effectively improves the accuracy in this regard.

Abstract:In endoscopic gastric polyp detection based on computer vision, efficiently extracting the features of images of small polyps is a difficulty in the design of a deep learning-based computer vision model. To solve this problem, this study proposes a detection model based on an improved you only look once version 4 (YOLOv4), namely YOLOv4-polyp. Specifically, on the basis of YOLOv4, this study adds a convolutional block attention module (CBAM) to enhance the feature extraction capability of the model in complex environments. Then, a lightweight CSPDarknet-49 network model is designed to both reduce the complexity of the model and improve its detection accuracy and detection speed. Finally, considering the characteristics of the gastric polyp datasets, the K-means++ clustering algorithm is used for the cluster analysis of the gastric polyp datasets and the attainment of the optimized anchor box. The experimental comparison results show that compared with the classical YOLOv4 model, the proposed YOLOv4-polyp achieves favorable detection performance on the two datasets as it improves the average detection accuracy by 5.21% and 2.05%, respectively, without compromising the detection speed.

Abstract:The problem of “alarm flooding” caused by the increase in the number of industrial alarm variables has seriously affected the proper functions of alarm systems. In response, this study proposes a method to learn the transfer relationship of alarm variables from the process alarm data set. First, leveraging the ability of transfer entropy to accurately measure the causal relationship between variables with first-order or multi-order autocorrelation, the study identifies the causal relationship between variables. Second, depending on the entropy value between variables, the nodes that transmit a large amount of information are reserved. Finally, the K2 algorithm is improved by considering the time proportions of variables in different states, and the final alarm propagation network is obtained by learning. The verification on the Tennessee Eastman process data set reveals that the method can judge the root cause of the alarm and well achieve alarm propagation paths through learning.

Abstract:Aiming at the problems of less target feature information and low contrast between foreground and background in satellite video, this study proposes a target tracking method integrating motion information and attention mechanism based on SiamCAR. First, the motion excitation and channel attention modules are introduced to enhance the target feature extraction information. Then, adjacent frames are regarded as new templates and added to the network to form a triple network and supplement template information. Finally, the Kalman filter algorithm is added to predict the target’s trajectory, and a prediction template is introduced to the network to construct a quadruple network and increase the motion information of the target. In addition, 10 sets of data in the SatSOT satellite video data set are selected for testing. The experimental results show that compared with those of the SiamCAR network, the tracking accuracy and success rate of the improved algorithm are increased by 6% and 6.2%, respectively.

Abstract:Aiming at the problems of the high dimension of features, high complexity of feature processing, and low efficiency of model detection of traditional industrial control network traffic data in complex network environments, this study uses an abnormal network flow identification and detection method based on random forest (RF) and long short-term memory (LSTM) network. Firstly, the random forest algorithm is used to calculate the importance score of flow characteristics, screen out important features, and eliminate redundant features. Then, LSTM is adopted to identify and detect abnormal flows. In order to evaluate the effectiveness and superiority of the model, the accuracy, precision, recall, and F1-score are used in this study to evaluate the model, and the model is compared with traditional machine learning methods including Naive Bayes, QDA, and KNN algorithms. The experimental results show that the overall accuracy of abnormal flow identification reaches 99% on the CIC-IDS-2017 public data set. In addition, compared with traditional machine learning algorithms, the proposed method has effectively improved the accuracy and efficiency of anomaly detection in complex network environments, and it has practical application value in industrial control network security and anomaly detection.

Abstract:Monotone composite quantile regression neural network (MCQRNN) cannot analyze the time series information and internal laws in load data well. In order to address this issue, this study combines MCQRNN and dilated causal convolutional networks (DCC) and proposes a new quantile regression model named, MCQRDCC. This model divides the input into quantile input and unconstrained input to make the output of the model increase with the increase in quantile, so as to solve the problem of quantile crossing. At the same time, the DCC structure is used to help the model fully analyze the sequence information in the load data and make the prediction results more in line with the changing trend of the real load. In addition, MCQRNN utilizes the exponential function to transform the constraint weight matrix and the hidden layer weight, which will affect the weight adjustment during backpropagation. In this study, the ReLU function is used instead of the exponential function to solve this problem and improve the prediction accuracy, and real load data is adopted for experiments. The experimental results show that MCQRDCC can effectively improve prediction accuracy. Compared with those of MCQRNN, the average Pinball loss and CWC of MCQRDCC are decreased by 2.11% and 9.31%, respectively, and AIS is increased by 10.51%.

Abstract:A lightweight YOLOv4 algorithm, MA-YOLOv4, is proposed to enhance feature fusion for addressing problems of complex detection networks, a large number of parameters, and poor real-time detection in deep learning-based maritime ship target detection tasks. Firstly, MobileNetv3 is employed to replace the backbone network, and a new activation function SiLU is introduced. The depthwise separable convolution is applied to replace the ordinary 3×3 convolution to reduce the number of network parameters. Secondly, an adaptive spatial feature fusion module is added to enhance feature fusion. Finally, the MDK-means clustering algorithm is adopted to get the anchor frame suitable for the ship target, and the Ship7000 dataset is utilized for training and evaluation. Experimental results show that compared with YOLOv4, the improved algorithm can reduce the number of model parameters by 82% and increase mAP by 2.57% and FPS by 30 f/s, which can achieve high-precision real-time detection of ships at sea.

Abstract:To effectively improve the transmission rate and reduce bandwidth burden, this study proposes a multiple-image encryption scheme based on compressed sensing and a hyperchaotic system. Specifically, several original images are spliced into a new plaintext image, and some plaintext information is combined with random positive integers to generate the initial value of the chaotic system. The pseudo-random sequence generated by the hyperchaotic system is utilized to produce the measurement matrix, scrambled sequence, and diffusion sequence the encryption process needs. Then, discrete wavelet transform, thresholding, and parallel measurement are performed to compress the plaintext image, which can effectively reduce the amount of operation data and greatly speed up the operation speed. Finally, the final ciphertext image is obtained by non-repeated scrambling and bidirectional mode-adding diffusion. Multiple levels of simulation experiments verify that the proposed algorithm can effectively resist cropping attacks and offer high security.

Abstract:As the inaccurate estimation of atmospheric light curtain and atmospheric light result in the halo effect, color distortion, and low contrast in the process of haze image restoration, a dehazing algorithm based on weighted least square (WLS) filtering and restorative controlling factor is proposed. Firstly, this study analyzes the principle and performance of the WLS filter, which can be utilized to effectively estimate the atmospheric light curtain. Secondly, with the assistance of the Sobel operator, the binary image edges are detected. The number of edges and the mean value of pixels are taken as the bases of the quad-tree space index, which improves the estimation accuracy of the atmospheric light. Finally, according to the causes of color distortion in the sky area, a restorative controlling factor is introduced to improve visual effects. Experimental results show that the mean gradient obtained by this method increases by 58.03%, and the information entropy increases by 2.88%. In particular, the running time relatively decreases by more than 50%. The proposed method achieves better restoration in terms of contrast, visibility, and color fidelity of the haze image containing complicated near and distant scenes that mixed dense haze, mist, and sky area.

Abstract:Road cracks are the key element of road damage, the classification of which can be used to arrange the formulation of road maintenance strategy in targeted ways. As the classification with manual annotation is time-consuming and inefficient, this study proposes a road-crack image classification method based on contrastive learning. In the traditional contrastive learning framework, the feature extraction part is improved to make the model more sensitive to the features of small cracks. Firstly, the data is augmented; the ResNet50 part is improved in feature extraction, and the multi-scale method is used to extract features. Then, multilayer perception (MLP) is employed to reduce the dimensions of extracted features and project them onto vector space. Finally, cosine similarity and cross-entropy loss of normalized temperature scale are applied to optimize the model. The experimental results reveal that compared with the original model, the improved model has a classification effect of 92.1% on crack images, an increase of 0.22%, which indicates it has a good effect on crack image classification.

Abstract:A core image stitching method based on Laplacian pyramid fusion with the best seam-line is proposed to address the problems of low core image stitching efficiency and the tendency of ghosting. Firstly, two core images to be stitched are processed through grey-level transformation, and then feature points are calculated and described according to the ORB algorithm. Secondly, the improved random sample consensus (RANSAC) algorithm is used to purify the feature points and complete feature point matching. According to the matched feature points, the alignment relationship between the images is calculated. Finally, the Laplacian pyramid fusion of the core images is realized based on the best seam-line, and the stitching is completed. The experimental results show that the improved RANSAC algorithm can improve the speed while ensuring accuracy, and the proposed image fusion method avoids the generation of ghosting and performs better on the PSNR, SSIM, and DoEM objective evaluation indexes of the fusion region compared with the other two image fusion algorithms.

Abstract:There are many kinds of natural disasters, and it is relatively difficult to semantically segment remote sensing images. In order to better realize remote sensing image segmentation, this study proposes a three-layer semantic segmentation model for remote sensing images based on a generative adversarial network. For the analysis of different scenes, a multi-level remote-sensing semantic segmentation framework is designed based on a fully convolutional network (FCN). The semantic segmentation of remote sensing images is effectively performed, and thus the segmentation accuracy of the model is enhanced. Experiments show that this model is effective, which can be directly observed from the segmentation results of damaged buildings, with mIoU being 82.28 %. In addition, this model is compared with other network models, and its performance evaluation index is significantly better than that of other network models. Finally, a reliable data report is provided to emergency management departments by analyzing various scene images of natural disasters.

Abstract:With the popularization of on-board GPS positioning equipment, a large amount of vehicle trajectory data and location information have been generated, and various trajectory mining technologies have emerged as the times require. However, the existing trajectory mining technologies rarely consider the leakage of users’ privacy. Therefore, this study proposes a method of stay point mining from vehicle trajectory data integrating privacy protection. In this algorithm, the stay points in the trajectory are screened out by density clustering, and privacy protection of the stay points is then conducted with the differential privacy technology. The experimental verification shows that the proposed method can not only effectively identify the location of the stay points but also protect their privacy from being leaked.

Abstract:A light guide plate (LGP) is the main component of the backlight module of a liquid crystal display (LCD), whose defects can directly affect the display effect of LCD. To address the problems of complex texture background, low contrast, and small defect size of LGP images, this study proposes an AYOLOv5s network for defect detection of large-size LGP images. First, the LGP image is divided into different images. Then, Transformer and the attention mechanism coordinate attention are integrated in the main part and feature fusion part, and the Meta-ACON activation function is selected. Finally, massive experiments are carried out on the basis of the self-built data set LGPDD. The experimental results indicate that the defect detection algorithm for LGP enjoys the mean average accuracy (mAP) of up to 99.20% and FPS of 77, which can realize good effects in the practical detection of bright spots, scratches, foreign bodies, bumps, dirt, and other defects in the 17-inch LGP in 12 s/pcs.

Abstract:For the severe challenges brought by the fluctuation and randomness of photovoltaic power generation to the load prediction of the dispatching department and the safe operation of the power grid, this study proposes a photovoltaic power prediction method of bidirectional long short-term memory (BiLSTM) optimized by variational modal decomposition (VMD) and cuckoo search (CS) algorithm. Firstly, VMD is employed to decompose the photovoltaic power sequence into sub-modes with different frequencies, and Pearson correlation analysis is adopted to determine the key meteorological factors affecting each mode. Secondly, the hybrid photovoltaic power prediction models of attention mechanism (AM) and BiLSTM are constructed, and the CS algorithm is utilized to obtain the optimal weight and threshold of the network. Finally, the prediction results of different modes are superimposed to obtain the final prediction results. The effectiveness of the proposed model is verified by predicting the output power of photovoltaic power stations in Arizona.

Abstract:Although direct-current (DC) charging piles are effective power supply equipment for electric vehicles (EVs), their frequent faults pose a threat to the charging safety of EVs. Accurately predicting charging pile faults can effectively ensure the safety of EVs in the charging process. For this reason, a fault prediction model for DC charging piles based on an improved gated recurrent unit (GRU) is proposed in this study. Specifically, the common fault types of DC charging piles during charging are analyzed. Considering the small sample size of specific fault data in the actual collection, variational autoencoder (VAE)-based data augmentation is performed to expand the sample data. Then, on the basis of the current fault prediction method based on the GRU network model, this study resorts to the particle swarm optimization (PSO) algorithm to optimize GRU network parameters, employs the support vector machine (SVM) model to improve the classification function output by the network, and thereby proposes a PSO-GRU-SVM fault diagnosis model for DC charging piles. Finally, an example is discussed to compare the prediction accuracy before and after the improvement, and the confusion matrix heatmaps are comparatively analyzed. Furthermore, the proposed model is compared with two commonly used network models. The results show that the proposed method can effectively improve prediction accuracy and thus verify the feasibility of the proposed method.

Abstract:In the field of information extraction, it is a basic and important task to extract entity relations from unstructured texts, and challenges such as entity overlap and model error accumulation often appear. This study is relation-oriented, and it proposes an improved joint extraction method for entity relations. The method divides the entity relation extraction task into two subtasks: relation extraction and entity extraction. For the relation extraction subtask, a self-attention mechanism is adopted to evaluate the degree of association between words, so as to simulate entity information and represent the whole sentence information by the average pooling. For the entity extraction subtask, according to relation information, the conditional random field is used to identify the entity pairs under the relation. This method can not only solve the problem of entity overlap by using the idea that relation and entity pairs coexist but also perform training by using the known relation in the dataset to make the entity extraction module independent from the results of the relation extraction module during the training, so as to avoid error accumulation. Finally, the effectiveness of the model is verified on the public datasets of WebNLG and NYT.

Abstract:The traditional CNN models have a poor weather classification effect for aerial video images and cannot satisfy the applications to mobile devices, and the existing weather image datasets are lacking, with single scenes. To address these problems, this study constructs four types of UAV aerial weather image datasets of sunny days, rainy days, snowy days, and foggy days for multiple scenes and proposes a weather scene classification model for UAV aerial video images based on lightweight transfer learning. The model uses a transfer learning approach to train two lightweight CNNs on the ImageNet dataset and designs three lightweight CNN branches for feature extraction. In feature extraction, EfficientNet-b0, a modification of the ECANet attention mechanism, is first used as the main branch to extract whole-image features, and two MobileNetv2 branches are employed to extract deep features unique to the sky and non-sky localities separately. Next, feature fusion is carried out for the three regions by Concatenate. Finally, a Softmax layer is used to classify the four classes of weather scenes. The experimental results indicate that the method achieves the accuracy of 97.3% in classifying weather scenes when applied to mobile and other computationally constrained devices, with good classification results.

Abstract:In order to make the model lightweight and facilitate the embedding of mobile devices, the YOLOv4 network is improved. Firstly, MobileNetV3 is used as the backbone network, and a deep separable convolution is adopted to replace the ordinary convolution of an enhanced feature extraction network, so as to reduce the number of model parameters. Secondly, when the feature map with a size of 104×104 is output, the dilated convolution with a dilated rate of 2 is fused, and it is then fused with a feature layer with a size of 52×52, so as to obtain more semantic and location information, which can refine the feature extraction ability and improve the detection performance of the model for minimal targets. Finally, the original pooling layer is connected in series with three Maxpools with a size of 5×5 to reduce the computational load and improve the detection speed. The experimental results show that on Huawei Cloud 2020 dataset, the mAP of the improved algorithm is improved by 2.33% compared with the YM algorithm, and on the public dataset VOC07 + 12, the mAP is improved by 3.12%, and the FPS has more than doubled compared with the original YOLOv4 algorithm, with the number of parameters reduced to 18% of the original one. As a result, the effectiveness of the improved algorithm is verified.

Abstract:In the power system, distribution scheduling is complex and well-coordinated, which mostly depends on the experience and subjective judgment of staff and is prone to mistakes. Therefore, it is urgent to use intelligent means to help analyze and generate maintenance plans. Named entity recognition is a key technology in the construction of the knowledge graph of power distribution networks and the question answering system, which can recognize named entities in unstructured data. In view of the complexity and strong correlation of distribution maintenance data, this study adopts the deep learning model BERT-IDCNN-BiLSM-CRF. Compared with the traditional model BERT-BiLSTM-CRF, this model integrates the neural network model IDCNN, makes better use of the performance of GPU, and improves the efficiency on the premise of ensuring recognition accuracy. The labeled maintenance plan data are trained, and the proposed model is compared with other commonly used models. The results reveal that the proposed model achieves the best effect in terms of the recall rate, accuracy rate, and F1 value, and its F1 value can reach 83.1%. The model has achieved good results in the recognition of distribution network data.

Abstract:To improve the efficiency of repairing faulty nodes in a distributed storage system, this study proposes a new construction algorithm for fractional repetition (FR) codes. The algorithm resorts to the factorization of the complete graph for node construction, and the nodes constructed are referred to as complete graph factorization based FR (CGFBFR) nodes. Specifically, the complete graph is factorized, and the number of factors generated from the complete graph after the factorization is completed is determined. The number of factors of the complete graph is selected according to the repetition degrees of the data blocks that need to be stored. All the vertices of the selected factors of the complete graph are regarded as the data blocks that need to be stored in the distributed storage system. Then, the edges of the selected factor graph are marked and stored as distributed data nodes. Finally, an encoding matrix is generated with the vertices and edges of the selected factors, and the distributed storage system stores the data blocks respectively according to the data in the encoding matrix. The experimental simulation results show that compared with the Reed-Solomon (RS) codes, the simple regenerating codes (SRCs), and the latest cyclic variable FR (VFR) codes in the distributed storage system, the codes generated by the new FR code construction algorithm proposed in this study can quickly repair the faulty node when the system repairs the node. The proposed algorithm can be widely applied to distributed storage systems as it reduces the repair bandwidth overhead, repair locality, and repair complexity of the faulty node, provides a simple construction process, and allows flexible selection of construction parameters.

Abstract:The dot product function is a first-level basic function in the BLAS library, which is widely called by scientific calculations and other fields. As the floating-point calculation introduces rounding errors, the double-precision dot product is unable to meet the accuracy requirements in some application fields, and thus high-precision algorithms are needed to achieve more accurate and reliable calculations. In this study, on the basis of the existing BLAS library, the interface of the high-precision dot product function is added to meet the high-precision requirements of applications on the domestic SW1621 platform. At the same time, the high-precision dot product algorithm uses such optimization strategies as loop expansion, visit-memory optimization, and instruction rearrangement to realize assembly-level manual optimization. The experimental results indicate that the high-precision dot product algorithm has the accuracy approximately twice that of the double-precision dot product, which effectively improves the precision of the original algorithm. On this basis, the average performance speedup of the high-precision dot product function reaches 1.61 after optimization.

Abstract:To address the problems of low contrast between target and background and lack of target feature information facing satellite video images, this study proposes a target segmentation and tracking method combining target motion information, spatio-temporal background, and appearance model. After the target area is obtained by positioning in the first frame, the histogram of oriented gradient method is employed to extract the features of the target, and the kernel correlation filter (KCF) is utilized to obtain the target tracking area 1. Subsequently, color and spatial features are used to build a spatial model of the context information about the target and its surrounding area and thereby obtain the target tracking area 2. Then, the visual background extraction algorithm is applied to detect the moving target in the target area in pixels and further obtain the segmentation area 3 of the single target. Finally, the correlation of the three areas is calculated, respectively, to obtain the optimal area as the final target tracking position and the template update sample. The experimental results show that compared with the KCF algorithm, the proposed algorithm obtains a significantly higher tracking success rate and accuracy and also achieves single target segmentation.