Research On Real-time Deep Neural Network For Image Semantic Segmentation

With the development of artificial intelligence technology,the proportion of intelligent devices in life is increasing day by day,and more and more human tasks start to be replaced by intelligent devices.Semantic segmentation,as one of the key issues in the field of computer vision,can realize the complete understanding of the scene by intelligent devices,which is the premise for better decision-making.At present,the research direction of semantic segmentation network focuses on the improvement of segmentation accuracy.The huge network scale and complex calculation process increase the difficulty of network model deployment,which makes the semantic segmentation meet the bottleneck in the process of life application.On the other hand,in the research direction of lightweight network semantic segmentation network,the current research results still have the problems of low real-time performance and poor balance performance.Aiming at the problem of low lightweight of semantic segmentation network,this dissertation proposes a lightweight algorithm model design method based on lightweight module design.By using lightweight convolution unit to replace the conventional convolution unit,the network lightweight is further improved.A lightweight network convolution module is built by using lightweight convolution units such as deep separable convolution,channel random mixing operation,residual link and affine operation.A lightweight semantic segmentation network model STNet is built by using multi-scale pyramid cascade structure and the lightweight network convolution module to construct network encoder and decoder.A lightweight semantic segmentation training algorithm based on knowledge distillation is proposed.By increasing the distillation channel and refining the learning granularity,the large-scale deep convolution neural network can guide the lightweight model to learn and train,so as to improve the accuracy of lightweight network semantic segmentation.In this dissertation,the semantic segmentation dataset Cityscapes is used to compare STNet and several novel lightweight semantic segmentation algorithms,and the hardware platform is built to realize the test and analysis in the actual scene.The experimental results show that STNet model has high real-time performance,great accuracy and great generalization performance.In the case of the same accuracy as other lightweight networks,the operation speed is increased by 160%.In the pure CPU hardware environment,the processing speed of1024×512 high-resolution image can reach 5fps,which can realize the real-time operation of low-power platform and achieve high segmentation precision in semantic segmentation task It has the effect of both low computation efficiency and high computation efficiency.At the same time,the semantic segmentation training algorithm designed in this dissertation improves the effect by 4.94% compared with the traditional training method,and improves the effect by 2.82% compared with the traditional knowledge distillation method.