Abstract:Skin color detection is widely used in many fields such as face detection, image filtering, and machine vision. There are many methods for skin color detection. This paper comprehensively reviews the pixel-based skin color detection methods, focusing on the principles and characteristics of statistics-based, threshold-based, and machine-based learning methods, and discusses their advantages and disadvantages. The development trend of pixel-based skin color detection methods is also discussed.

Abstract:Cloud computing is one of the emerging industries based on the Internet for commercial calculation model. It provides a quick and easy and reliable access to network resources. Cloud computing is introduced. The task scheduling in cloud computing is analyzed, and the research status of cloud computing task scheduling algorithm are classified and summarized according to different scheduling goal. The task scheduling algorithm can be divided into single objective optimization algorithm and multi-objective task scheduling algorithm. The representative algorithms of each method are analyzed, and the advantages and disadvantages of each algorithm are compared and summarized in detail, and the way of improvement is also inducted.

Abstract:As a new generation of high-precision biometric recognition technology, palm vein pattern recognition technology is widely used in the field of personal identification. However, its recognition effect is limited by the quality of the image. Low-quality images often result in low recognition accuracy. How to effectively evaluate the image quality and screen out high-quality images becomes an important research issue in palm vein recognition technology. This study aims to solve this problem and proposes a multi-parameter palm vein image quality evaluation method based on BP-AdaBoost neural network. According to the quality characteristics of the palm vein image, the evaluation indexes (contrast, entropy, sharpness and equivalent visual number (enl)) of multiple parameters are proposed. Based on the excellent nonlinear fitting characteristics of BP network, multiple evaluation parameters are used as network input, the classification result is network output, and 10 BP weak classifiers are trained. On this basis, the final strong classifier is obtained by AdaBoost algorithm. The experimental results show that compared with the traditional weighted fusion evaluation classification method, the classification results have higher accuracy and the system has good application value.

Abstract:With the continuous development of English and Chinese speech recognition technology, the research on minority language speech recognition technology has followed closely and achieved certain results. The Tibetan people are an indispensable member of the Chinese nation's family. The study of Tibetan speech recognition technology is an indispensable part of the research of speech recognition technology. Firstly, the paper presents the research process and research improvement of Tibetan speech recognition in China. Secondly, it introduces the template-based matching and statistical probability model and artificial neural network used in Tibetan speech recognition research from the characteristics of Tibetan language itself and its pronunciation features and elements, then summarizes the characteristics and application scope of the three methods. Finally, it discusses the research progress of Tibetan speech recognition and the characteristics of each recognition method, discusses the difficult problem and the direction of its future development.

Abstract:Prenatal physical examination of pregnant women is a import part of perinatal medicine. Prenatal prediction of fetal weight can provide an accurate reference for judging the healthy development of the fetus. The multiple physical examination records have the characteristics of variable time interval distribution during gestational period. This study proposes a variant of LSTM model, Variable Time Interval LSTM (VTI-LSTM), to solve the variable time intervals problem. The data of this study were from 122 462 medical records of 10 473 pregnant women from several women's hospitals during 2015 to 2018. The experiments of fetal weight prediction compare the traditional formula estimation methods with the machine learning methods such as GBDT, MLP, SVR, RNN, LSTM, and VTI-LSTM. The results show that Variable Time Interval LSTM has a good prediction result in the prediction of low birth-weight fetal and macrosomia.

Abstract:Considering the green scheduling problem of TFT-LCD manufacturing cell stage based on improved cuckoo search algorithm, a mathematical model was established aiming at minimization of the maximum completion time and total carbon emissions. By using three-stage coding based on machine selection, speed selection and process selection and using an improved cuckoo search algorithm with dynamic coefficients before step size factor, the Pareto optimal solution set is constructed by combining the dual championship and the dynamic elimination system. The validity of the model and algorithm is verified by simulating the actual production data of a workshop. The simulation results show that the improved cuckoo search algorithm can effectively reduce carbon emissions while guaranteeing the maximum completion time.

Abstract:Lane change plays a vital role in traffic safety. Accurately predicting the driver's lane change behavior can significantly improve driving safety. In this study, a hybrid neural network based on fully connected neural network and recurrent neural network is proposed to accurately predict lane change behavior. And a dynamic time window is proposed to extract lane change features including driver physiological data and vehicle kinematics data. Finally, the effectiveness of the proposed model is verified by the data in real traffic scenarios. In addition, the proposed model is compared with five other predictive models. The results show that the proposed model has higher accuracy and perspective time than other models.

Abstract:Big data technology plays an increasingly important role in analyzing and mining traffic big data. In order to quickly and effectively analyze the operating mode and passenger carrying strategy of taxis, this study designed the effectiveness index model to quantificate and sort the taxis' effectiveness. Taking high-effective taxis as the research object, a data processing and visualization platform is developed based on Spark big data framework. Firstly, high-effective taxis trajectory data are processed to obtain characteristic data for visualization. Then visual analysis is carried out, including high-effective taxis operation characteristics obtained from statistical analysis and interactive chart display, using hexagon grid and DBSCAN algorithm to visualize the hotspot of high-effective taxis carrying passenger points in different time periods, implementing interactive trajectory query based on buffer, and extracting the trajectory-related factor. Finally, the validity and reliability of this platform are verified by GPS trajectory data of Chengdu taxi.

Abstract:This study applies facial recognition, human eye recognition, and speech recognition technology to consumer population analysis, and proposes a system that can collect consumer data in multiple dimensions and conduct intelligent product recommendation. Different from traditional data collection method, this system collects implicit evaluation data and collects explicit evaluation data in the meantime, which can effectively increase credibility of collected data. The system exploits on facial recognition to obtain consumers' facial features, makes use of human eye recognition to track down duration of human eyes staying and gazing at products, uses speech recognition to obtain sentiment polarity of texts and evaluation keywords, and uses a recommendation model based on user face attributes to recommend products. Through the experiments, we discover that 80% of the experimenters express their satisfaction with the first five recommended products introduced by the system, which was formerly trained by 20 other experimenters, showing that multi-dimensional collection of consumer data can break the limits of traditional data collection and possesses a higher credibility.

Abstract:At present, the calibration of water meters in China mostly adopts manual calibration, and there are many repetitive manual operations in the process, which makes the calibration process time-consuming and laborious. In order to solve this problem, robots were used to replace the manual work to complete the calibration in the process of water meter calibration and a water meter grasping method based on machine vision was proposed in this study. Different types of water meters in different environments can be detected through the YOLOv3 network in the system. After obtaining the type and position of the target water meter, the system carries on the dection of position and posture of the water meter to obtain the coordinates of the grasping point and the attitude angle of the water meter, and controls the robot to grasp the water meter. Experimental results show that different types of water meter in different external environments can be grasped and accurately placed with the system, the system has sound robustness and high grasping success rate, and can satisfy the needs of actual water meter automatic calibration line.

Abstract:In view of the safety hazard of passengers falling down on the escalator, a machine-based human fall behavior recognition system and an escalator automatic emergency stop device were designed. The OpenPose human joint detection algorithm is used to extract the bone characteristics of the human body. The Inception V3 network model is used to build the classifier, and the collected bone feature information is classified to identify the fall behavior of passenger. The training results show that the test accuracy of single and multi-person samples is up to 98.9% and 80.0%. After the fall behavior is identified, the test results are wirelessly transmitted to the emergency stop device based on the STM32 microcontroller and various sensors. Finally, the experimental test is carried out in the simulated escalator environment. The test results show that the control of the escalator automatic emergency stop system has good real-time performance.

Abstract:Computer vision technology is widely used in autopilot system, which mainly solves the problem of object recognition and object classification. In this study, a lightweight neural network structure is proposed according to the task. In order to solve the problem of insufficient training data, an improved data enhancement algorithm is adopted to double the training data. At the same time, in order to solve the problem of using data generator as verification set and unable to use tensorboard, a solution is proposed. The principle of neural network processing image information is studied in detail by convolution network visualization method, and the optimization method is put forward. The accuracy of the trained model is 97.5% on the verification set, which meets the accuracy needs of autopilot system for classification task.

Abstract:The optimized Mask R-CNN network based on deep learning is used to visual detection of the tiny defects on gears. Firstly, by comparing the detection effects of five kinds of residual neural network, resnet-101 is selected as the image sharing feature extraction network. Then, the detection rate for missing tooth is correspondingly improved by eliminating the unreasonable 3×3 convolution of feature map P₅ in the feature pyramid network. Finally, in order to effectively train the region proposal network, the appropriate anchor size and aspect ratio are set according to small fluctuation of annotation box in the designed sample labeling scheme. The optimized Mask R-CNN network eventually achieved 98.2% detection rate for missing tooth on gears.

Abstract:In recent years, with the development of embedded computers as well as the appearance of various kinds of heterogeneous embedded hardware, it is a trend to improve the portability of operating systems and the code reusability of application programs. In this paper, we introduce a platform abstraction layer with strong universality, which is designed for Linux and ReWorks operating systems and hardware platforms. Specially, the platform abstraction layer redesigns the application programming interface to provide a unified interface to users to develop various embedded applications. It is proved that the platform abstraction layer can improve the portability of operating systems and the code reusability of application programs. At the same time, it has reliability in terms of real-time performance.

Abstract:In view of the current intelligent mobile robot in the tracking process due to the target shape on the changes in a loss of tracking target, using the Caffe deep learning framework and ROS robot operating system as a development platform, a high accuracy and high real-time target tracking system of mobile robots is designed for research. The GOTURN target tracking algorithm based on the twin convolutional neural network, which is robust to target deformation, viewing angle, slight occlusion and illumination changes is used, and the ROS system is used as a bridge to enable the offline training tracking model to be applied to the TurtleBot mobile robot in real time, also a detailed test is carried out. Experimental results show that the target tracking system is not only feasible in design, but also has the characteristics of low cost, high performance and easy expansion.

Abstract:The processing failure information affects the real-time message processing of normal users and reduces the overall transmission performance of National Meteorological Communication System. This study refers to stateful message queue technology, introduces the transceiver status file, which monitored and dispatched all kinds of abnormal information in the message queue. The communication system independently creates data reprocessing and data redistribution processes to process abnormal messages according to the sleep time strategy and priority. It guarantees the fast processing of normal message queues to satisfy the requirement of the system reliability and transmission timeliness. After optimization by this scheme, the system runs stably in the abnormal state, and the transmission time of each time is obviously improved. The reliability and transmission performance of CTS are strengthened.

Abstract:Alarm processing is fundamental function in Telematics System. AJAX pulling is common method in traditional system. But AJAX pulling is not real-time communication, and the server suffers unnecessary requests. In this study, we design a telematics alarm pushing system based on Websocket and Redis. Then, we test this system. The results show that Websocket is an ideal technology for alarm pushing. It improves network throughput, reduces latency and network traffic.

Abstract:The port amount of import and export goods can reflect the congestion of port flow, whose accurate prediction would provide suggestions for port management to make reasonable decisions. In this study, the Seq2Seq model in the field of machine translation is used to model various factors that affect the amount of goods inflow and outflow from the port. An Seq2Seq model can reflect the change of the amount of import and export goods in the time dimension and describe the influence of external factors such as weather and holidays, so as to make accurate predictions. An Seq2Seq model consists of two LSTM, respectively acting as an encoder and a decoder. It can capture the changing trend of containers in the short and long term and predict the amount of goods in the future based on historical import and export volume. Experiments were carried out on a real-world dataset of import and export containers in Tianjin Port. The experimental result reveals that the deep learning prediction model based on Seq2Seq is more effective and efficient than traditional time series model as well as other existing machine learning prediction models.

Abstract:In the edge detection model based on neural network, most of models have lower detection efficiency, and the detection effect needs to be improved. Inspired by the characteristics of biological vision systems, a new edge detection method based on Gaussian Positive-Negative (GPN) radial basis neural network is proposed in this study. Firstly, we construct a new GPN radial-based neural network, which takes each pixel point preprocessed by Gaussian filtering in the image as the center point of the GPN radial-based neural network and inputs it into the neural network. Then, the partial characteristics of the convolutional neural network are used between each layer for processing, and the results are output after the expansion layer and the hidden layer are calculated. Finally, the edges are extracted by the contour tracking method according to the output result. In this study, the corresponding numerical experiments are carried out in two aspects:detection effect and efficiency. For the composite images and some intensity inhomogeneous images, compared with the pulse-coupled neural network model, the genetic neural network model and the convolutional neural network model, the proposed model is improved in efficiency and the edge connectivity is better. The experimental results show that the proposed edge detection method based on GPN radial basis neural network is a new and effective edge detection method, which is more efficient than the traditional neural network edge detection method, and the detection effect is also improved.

Abstract:Postpartum hemorrhage in pregnant women is one of the most important factors of maternal death around the world, ranking first in China. However, the early diagnosis of postpartum hemorrhage has always been a medical problem. With the popularity of Electronic Health Records and the development of machine learning and deep learning technologies, new solutions have been provided for predicting postpartum hemorrhage in pregnant women. This study proposes to construct a mixed prediction model of postpartum hemorrhage based on LSTM and XGBoost by using the Electronic Health Records of pregnant women. The experimental results show that the hybrid model based on LSTM and XGBoost is feasible to predict postpartum hemorrhage in pregnant women. It can provide a reference for doctors to judge the situation of postpartum hemorrhage and provide decision-making support for whether blood preparation would be needed during delivery. It is of positive significance to reduce the mortality rate of postpartum hemorrhage.

Abstract:Aiming at the problem that many documents categorization methods seldom control hybrid variables and have low robustness to data distribution, a documents (text) categorization method based on covariate adjustment is proposed. Firstly, it is assumed that the confounding factors (variables) in text categorization can be observed in the training stage, but not in the testing stage. Then, the sum of confounding factors is calculated in the prediction stage under the condition of the confounding factors in the training stage. Finally, based on Pearl's covariate adjustment, the accuracy of text features and classification variables to the classifier is observed by controlling the confounding factors. The performance of the proposed method is verified by microblog data set and IMDB data set. The experimental results show that the proposed method can achieve higher classification accuracy and robustness against mixed variables than other methods.

Abstract:With the development of smart grid technology, short-term power load forecasting becomes more and more important. To improve the accuracy of short-term electric load forecasting for individual users, this study proposes a load forecasting model, which is based on K-means and Convolutional Neural Network (CNN). Firstly, K-means is applied to group users into two categories. For users with strong daily correlation, the historical loads of adjacent time points and same time points in adjacent days are taken as input to the CNN model to extract abstract features for prediction. For users with weak daily correlation, the historical loads of the adjacent time points are utilized as features. To assess the performance of the proposed method, we conducted comparison experiments on real data with random forest and support vector regression. The experimental results show that the MAPE of the proposed approach is reduced by 20%.

Abstract:Aiming at the inaccuracy of English teaching quality evaluation, a teaching quality evaluation method based on Genetic Algorithm (GA) to optimize RBF neural network is proposed. Firstly, the principal component analysis is used to select the evaluation index of teaching quality, then the RBF neural network teaching evaluation model is designed, and the initial weight of RBF neural network is optimized by GA. The experimental results show that the method can effectively evaluate the quality of English teaching, and has high accuracy and real-time.

Abstract:Aiming at the most important motion characteristics of human behavior, a second-level recursive anomaly behavior recognition method based on time context is proposed. Different from traditional deep learning training methods, this method does not directly learn features from image data, but extracts them. The shape information HOG feature is used as the training input. Firstly, the image shape feature based on the HOG algorithm is extracted, and the extracted feature is used to train the DBN network. Secondly, the trained DBN network and the Softmax classifier are used to identify the human body coarse target region, and then according to the coarse The time-series context information of the target area, calculate the centroid acceleration. Finally, the threshold of the acceleration is judged, and the precise target area of the abnormal behavior is identified. This paper applies the two-level recursive method combining the weight and the target to the classroom behavior recognition, and the experimental results show that the The method can better recognize the classroom behavior in the scenes of motion blur and target dense occlusion, and the recognition rate is greatly improved compared with other methods. Classroom abnormal behavior data analysis can play a supporting role in classroom dynamic management and learning effect evaluation.

Abstract:When using the current algorithm to detect the discrete data in the medical database, problems such as long execution time, low detection efficiency and low detection rate of discrete points are caused by the lack of data filtering and other processes. Therefore, an algorithm for detecting discrete data in the medical database based on hierarchical deep learning is proposed. Firstly, the dynamic grid method is used to divide the sparse and dense areas in the space, so as to reduce the size of data detection and shorten the detection execution time. Then, the expert knowledge and data attribute value distribution information are integrated through the hierarchical deep learning process, and realize the detection of discrete data in medical database. Experimental results show that this algorithm can accurately complete the detection of discrete data in the medical database in a relatively short time, and has more advantages in application compared with the traditional algorithm.

Abstract:Considering the conflicts between vehicles and between vehicle and pedestrian at the intersection with signal control, the microscopic traffic simulation model of the intersection is constructed under the Discrete Event System Specification (DEVS). By calibrating simulation parameters with observation data of a typical intersection in a city, the simulation results are compared with the calculated traffic capacity according to the "Specification for Urban Road Engineering", and the model is verified. On this basis, first, the influence of the amount of left-turn ratio on the traffic capacity of intersection is analyzed by the simulation. Then, an intelligent green ratio control strategy is designed based on the number of vehicles waiting to cross the intersection in each direction. The simulation results show that, the capacity increases first and then decreases with the increase of the ratio of left-turn vehicles; and the intelligent green ratio control can significantly improve the traffic capacity of intersection and significantly reduce the average approach road delay time. These prove that the simulation model can truly simulate the interaction of various intersection factors, and is easy to expand and universalize, which can be applied to the study of other intelligent traffic problems.

Abstract:Aiming at the research of image copying and tampering detection and tampering area localization, a tamper image detection algorithm for low frequency fast Tchebichef moment is proposed. Firstly, the image is decomposed by non-sampling wavelet transform, and the low-frequency part of the image is selected for overlapping segmentation. The improved low frequency fast Tchebichef moment is extracted as the feature vector. Then the PatchMatch algorithm is used to match the extracted block features. Finally, the dense linear fitting algorithm is used to remove the mismatch and the morphological operation is used to complete the final tamper region localization. Compared with the existing tampering image detection algorithm, the proposed algorithm has better positioning effect for single-region tampering, multiple tampering and multi-region tampering, and reduces the running time of the algorithm, and improves the real-time performance.

Abstract:Advanced SAT solvers solve large application instances with efficient branching heuristics. At present, the VSIDS strategy is the most representative branching strategy based on conflict analysis. It is widely used because of its robustness. However, in each conflict analysis, the incremental method of determining the variable activity is too single. To solve this problem, this study proposes a branch heuristic algorithm based on variable mixing features. The purpose is to fully utilize the different information features carried in the variables involved in conflict analysis to distinguish variables, and further guide the variable activity growth. The proposed branching strategy algorithm is embedded into Glucose4.1 to form the solver Glucose4.1+MFBS. Through experimental comparison and analysis, the experimental results show that the improved branching algorithm has certain advantages over the original VSIDS strategy, and the number of solutions increases obviously.

Abstract:Eye-movement interaction has a broad application prospect in the field of human-computer interaction. Aiming at the problems of traditional eye-movement interaction sensors, such as universal intrusiveness, complex calibration process and high price, low resolution of common monocular camera sensors, etc., an eye movement recognition method based on front-facing camera video using directional gradient histogram (HOG) features + SVM + LSTM neural network, and a simple human-computer interaction application are proposed in this study. Firstly, the region of eyes are localized and tracked after face alignment. Secondly, the open-close and non-blinking state of the eyes is judged by the SVM model. Then, the position of eye center between adjacent frames is analyzed to roughly judge the eye movements, and the suspicious interframe difference video sequence of intentional eye position is obtained, which is the input of the LSTM network for prediction, and then trigger computer commands to complete the interaction. Through the self-made data sample set (about 10% of which are negative samples), the accuracy of dynamic blink recognition is better than 95%, and the accuracy of eye movement behavior prediction is 99.3%.

Abstract:For multi-cluster problems, PSO-FCM cluster algorithm is lack of performance and easily leads to local optimum, which affects the accuracy of multi-cluster result. To tackle these issues, an optimized PSO-FCM cluster algorithm based on PCA is put forward. By introducing PCA processing method, setting different movement weight on each dimension of particle and reducing particle sensitivity, reasonably controlling movement speed of particles on each dimension and effectively decreasing unconstrained particles on each dimension, possibility of moving into false cluster is increased due to over-sensitive particles on interface of multi-cluster groups. This paper introduces related conditions of PSO-FCM algorithm briefly and the proposed optimized algorithm in detail. Finally, this paper presents the experiment results, i.e., the optimized algorithm proposed in this study is totally better than other algorithms in many data sets.

Abstract:YARN is a distributed resource management system of Hadoop. It can be used to improve the utilization of memory, I/O, network, disk and other resources of distributed cluster. However, there are many configuration parameters in YARN. Due to this reason, manual tuning of Hadoop performance to get the best performance is difficult and time-consuming. Based on the existing YARN resource scheduler, a successive approximation closed-loop feedback control method is proposed. This method can dynamically tune the parallel number of MapReduce (MR) jobs in the running state of the cluster, and eliminating the process of manual adjustment of parameters. Experiments show that the proposed approach reduces the MR operation time for 53% and 14% based on capacity scheduler and fair scheduler, respectively, compared with the default configuration.

Abstract:Given the limitations of the four-band images' color cast correction algorithm based on ternary linear polynomial regression, to better solve the RGBIR four-band image color cast problem of infrared crosstalk, three aspects, i.e., the sample, data type, and correction model of the polynomial regression algorithm, are worked on to improve the color cast correction effect of four-band image. The correction model was established by adding training samples of the algorithm and converting the data into signed floating-point pixel values to make the correction algorithm more robust. Due to the nonlinear gray-scale expression of RGB images, the ternary linear model was changed to the ternary quadratic one. Experiments show that the proposed optimization method in this study improves the color cast correction effect of the four-band images.

Abstract:As the high signaling cost and low data transmission efficiency of traditional mobile management strategy in 5G ultra-dense network deployment, this study proposed a handover algorithms in 5G based on SDN architecture. By the use of the global network state information in SDN controller to make the optimal handover decision, the delay and network resource cost caused by mobile node collecting network status information can be reduced. The proposed handover management strategy is simulated by Matlab. Compared with the handover management mechanism of LTE system, it has advantages in handover delay and average handover times.

Abstract:With the development of Internet and HTML5 technology, the application of file upload on the Web is more and more popular. At the same time, the requirement of efficiency, stability, security and universality of file upload, especially large file upload, is higher and higher. However, most of the current file upload methods can not meet these requirements very well. Therefore, this study proposes to use file fragmentation and multi-concurrent transmission method to improve the efficiency of large file transmission, so as to solve the related problems of multi-concurrent transmission control in the front end of the Web and repeated upload of the same file.

Abstract:In the complex traffic scene image, road segmentation is difficult and the edges of the segmentation are rough. In order to solve this problem, a road segmentation method based on multi-feature fusion and conditional random field is proposed. Firstly, the textons and color features of the image are extracted from the traffic image. Then, the road segmentation problem is regarded as a pixel-based binary classification problem. The extracted texton features and color features are fused and input into the SVM classifier, which can achieve the coarse segmentation of the road area and the background area in the traffic image. Finally, by using the color and position constraints of the fully connected conditional random field to optimize segmentation results, a smoother segmentation edge can be obtained and compared with other segmentation algorithms. The experimental results demonstrate that road segmentation method that based on the multi-feature fusion and the conditional random field achieves 95.37% of average segmentation accuracy and 94.55% of mean pixel accuracy.

Abstract:The multi-dimensional MEMS sensor is applied to elevator monitoring. According to the working characteristics of the elevator, the quaternion complementary filtering method is modified to correct the gyroscopedata to solve the real-time attitude of the elevator, and then the Kalman filtering method is applied to further improve the attitude monitoring accuracy. The actual verification shows that the method can improve the accuracy of elevator attitude monitoring data,and use the attitude angle and acceleration kurtosis to analyze and compare, which can provide critical data basis for elevator safety and comfort assessment.

Abstract:Topic island on Internet Web pages has seriously affected the performance of focused crawlers. The metric of setting more initial links to find new topics cannot guarantee the comprehensiveness of Web pages. On the basis of analyzing typical crawling strategies and taking into account the hierarchy of topic relevant, we propose a crawling strategy using dynamic tunneling. The crawling strategy uses the tunneling technology based on the topic of Web pages to discover new topics, and constructs a hierarchical topic model to solve the problem of weak link between two topic islands. Meanwhile, the strategy can effectively prevent topic drift caused by collecting too many topic-independent pages, thus dynamic controls the tunneling depth in the crawling direction with the semantic information of the topic maintained. Experimental results show that the proposed method can better address the topic island issue, thereby enhancing the recall of focused search engines.

Abstract:In the process of flower conservation, due to the lack of professional maintenance technology, the difficulty of flower conservation and high cost, this study proposes a method based on deep learning image classification technology to realize the full automatic flower maintenance. Because the growth of flowers is often affected by many factors, it is easy to make a wrong judgment on the growth of flowers by relying on the classification of flower growth status images. Therefore, this method designs two kinds of flower image features and growth environment parameters. The convolutional neural network of the input channel automatically recognizes the state of flower growth. Experiments show that this method can improve the recognition accuracy of flower growth status, and thus improve the level of automatic maintenance technology of flowers.

Abstract:Base on the rapid development of intelligent transportation, this work studies lane departure detection and vehicle forward safety distance detection technology under high-speed section based on the machine vision. First fix the car camera, obtain the camera's internal and external parameters through camera calibration, and then design the distance detection model which can not only detect the distance between the front vehicle and the unmanned vehicle, but also calculate the deflection angle of the front vehicle relative to the optical axis of the camera. Secondly, based on the CCP (The Car's Current Position) deviation detection algorithm, the safety and alarm zones are set to establish lane departure models, and the algorithm judges whether the current vehicle is deviated or not. Finally, the algorithm is transplanted to the embedded platform DSP-DM3730 by TI's DVSDK(Digital Video SDK). Experiments show that the vehicle distance detection model and lane departure model designed in this work have good reference value in solving the problems of forward collision detection and lane departure detection of unmanned vehicles.

Abstract:At present, many bearing production lines use manual naked eyes to identify the bearing workpiece number, which not only has poor recognition effect but also has low efficiency. In this study, a bearing embossing character recognition algorithm based on machine vision is designed, which is beneficial to bearing production and subsequent management. Firstly, the noise of the collected image is reduced by Gaussian filtering to reduce the influence of noise on the subsequent operation, and then the least square method is used to extract the ROI ring to determine the image region to be operated. Then the 1/8 circle scanning method is used to expand the ring image to make the character recognition operation more concise; then the character is segmented and normalized; finally, the character is recognized by SVM. The experimental results show that this method can realize bearing imprint character recognition, the recognition accuracy is more than 98%, and has good robustness, the system response speed is fast, and can meet the needs of industry.