Image semantic segmentation based on improved DeepLab V3 model

Abstract

To improve the image-segmentation speed based on the accuracy of a convolution neural network model, an improved DeepLab V3 network is proposed in this paper. The original feature extractor of DeepLab V3 is replaced with the lightweight network structure of MobileNet V2, and the original nonlinear activation function of a rectified linear unit is partially displaced by a new Swish activation function. Experimental results show that the improved DeepLab V3 network model can balance the segmentation accuracy and speed of the model better than the V3+ algorithm, which is the most accurate DeepLab network model till now. The running speed is improved significantly with a certain level of accuracy. In tests using different datasets, the running time decreased by 84% and 88.9%, and the model memory consumption decreased by approximately 96.6%. The improved DeepLab V3 network can adapt to deep-learning applications and satisfy their high-speed requirements.