Abstract:With the development of deep learning, many complex problems in semantic segmentation tasks are solved, which lays a solid foundation for image understanding. The proposed algorithm highlights two aspects. Firstly, our algorithm fuses multi-scale features from different levels of deep convolutional network by using multi-level deconvolution network. Then our algorithm upsamples these feature maps by deconvolution, meanwhile zooms them up to the original image size to predict semantic categories pixel-to-pixel. The second one, we propose a new method for data processing which is batch centralization algorithm, in order to improve the performance of network structure in this study. Through experimental verification, the mean IoU of semantic segmentation on the SIFT-Flow dataset reaches 45.2%, and the accuracy of geometric segmentation reaches 96.8%. The mean IoU of semantic segmentation on the PASCAL VOC2012 dataset reaches 73.5%.