Application Of Improved MobileNetV3 In Image Semantic Segmentation

Image semantic segmentation has been widely used in automatic driving,mobile robots and other fields.It can help drivers better analyze road conditions and surrounding information,and also help robots obtain information in the environment to build semantic maps and analyze scene information.This paper proposes a semantic segmentation backbone network with better segmentation performance and less computation: improved mobilenetv3.The specific work is as follows:(1)Through the analysis of the network structure and each module of mobilenetv3,it is found that the SE attention mechanism used by mobilenetv3 contains two full connection layers and has a large amount of computation.After comparing the lightweight CBAM and ECA attention mechanisms,the ECA attention mechanism is used to improve mobilenetv3.The parameter quantity is reduced by about 43%.(2)In order to improve the computational efficiency of sigmoid activation function in ECA attention mechanism,a linear approximate hard-sigmoid activation function is proposed to replace it,and a HECA attention mechanism is constructed.The HECA attention mechanism is used to replace the SE attention mechanism of mobilenetv3,which reduces the number of parameters by 43%.At the same time,the segmentation accuracy is improved by 0.7% on cityscapes dataset and 0.1% on VOC2012 Aug dataset.(3)The improved mobilenetv3 is used to replace Res Net as the backbone network of different semantic segmentation networks for comparison.Experiments show that using the improved mobilenetv3 backbone network can effectively reduce flops,params and GPU memory,and reduce 58.1% of params and 86.7% of flops in deeplabv3.In deeplabv3+,61.9% of params and 59.2% of flops are reduced.Deeplav3+uses less GPU memory and has better segmentation results after replacing the backbone network.