Accurate And Efficient Semantic Segmentation Algorithm Based On Separable Residual Modules

Semantic segmentation is an important part of scene parsing in Intelligent Vehicles(IV).Currently,the research of semantic segmentation algorithms is mainly divided into two aspects.One is constructing complex networks,so the accuracy can be improved at the expense of the efficiency.The other is constructing light networks and the efficiency can be improved at the expense of the accuracy.To solve the problem that current approaches cannot meet the demands on accuracy and efficiency simultaneously,in this paper,we propose an accurate and efficient algorithm for semantic segmentation.Residual module can construct a deep network without the problem of gradient vanishing and degradation,thus improving the accuracy of the network.However,the computational efficiency of traditional residual modules is an important factor affecting network efficiency.We redesign a new residual module,named as separable residual module,by combining depthwise separable convolution with residual connection,which is aimed to break the limitation in computational efficiency of the traditional residual module.The separable residual module inherits the advantages of the residual connection and the depthwise separable convolution.It can keep the number of parameters small while the network going deeper,which results in high efficiency for the residual module.In order to further improve the network efficiency,we propose a new downsampling module containing a convolution layer with stride 2 and a Max-Pooling layer with stride 2 in parallel.The downsampling module uses pooling operation to improve the computational efficiency of the network and uses the large receptive fields of convolution kernels to improve the accuracy of the network,thus achieving the balance between the efficiency and the accuracy.Based on the separable residual module and the parallel downsampling module,we propose a real-time and accurate semantic segmentation network.On NVIDIA TITAN X,a comprehensive set of experiments on Cityscapes dataset shows that the segmentation accuracy can reach 67.86% with the efficiency of 12 frames per second.The result demonstrates that the proposed method can achieve a good performance both in accuracy and efficiency compared with other state-of-the-art methods.It can be further applied to intelligent vision applications such as intelligent driving scene understanding,which need to maintain the high accuracy and real-time performance of the algorithm.