Research On Video Semantic Segmentation Based On Deep Convolution Neural Networks

At present,the field of automatic driving has made breakthroughs in feasibility and practicality,which has a far-reaching impact on the automotive industry and even the transportation industry.Image information can be obtained by using visual sensors,so image semantics segmentation is one of the most important basic problems in computer vision.Generally speaking,the process of image and video signal acquisition by visual sensors requires high efficiency,because the amount of video data is very large,while the traditional image processing methods are slow and has low accurate in real-time video semantics segmentation.Based on this,the improved U-shaped network and W-shaped network based on the deep convolution neural network are proposed in this paper,and the optical flow field is used to propagate and fuse the features between frames.Thererfore the whole method can reduce the processing time of real-time video semantics segmentation as much as possible while maintaining the fine segmentation accuracy.The main contents of this paper include:Firstly,based on the encoder-decoder structure,a U-shape-S-A network structure is proposed,which can recover spatial objects.The network structure uses the deep separable convolution structure and the channel attention model to remove the spatial and channel correlation in the network feature graph,and distinguishes the different contributions of different channels to the spatial detail recovery of the final semantics segmentation and the classification of the whole object category,which imporoved the classification accuracy furthermore.In order to reduce the processing time of segmentation,a W-shaped convolution neural network model based on ResNet network and Xception network is proposed.The whole model expands the single branch of U-shaped convolution neural network into two branches,and carries out fast down-sampling and maintains large-scale sensing field respectively,which can optimize the overall sensing field and detail information of the image at the same time.In the next,for improving the accuracy of segmentation,a feature propagation algorithm based on optical flow field is adopted.This algorithm can effectively utilize the correlation between frames.The previous frame in video is extracted and semantically segmented through the whole W-shaped network,and the final segmentation results are obtained.The latter frame not only extracts deep features through the whole W-shaped network,but also extracts deep features through the whole W-shaped network.Using the deep features of the previous frame propagating through the optical flow field,the corresponding features are aggregated,and then the semantics of the next step is segmented to get the experimental results.Compared with the previous W-shaped network,this method can further improve the detection accuracy of video semantics segmentation after feature aggregation through optical flow field.