DeepLabv3+ Semantic Segmentation Network Optimized by Double Weighted Feature Fusion
Proceedings of WCSE 2022 Spring Event: 2022 9th International Conference on Industrial Engineering and Applications · 2022
Intelligent identificat ion to prohibited items in lug-gage is of great significance to ensure the safety of passengers. In this paper, we propose CWF-Deep Labv3+ semantic segmentation network to improve the segmentation accuracy to small size objects in X-Ray security images, and reduce the object pixel loss caused by complex background. The CWF-Deep Labv3+ network is based on the DeepLabv3+ network, which improves both the Xception backbone and the Atrous Spatial Pyramid Pooling (ASPP) module. Firstly, we co mbine the Cross Stage Partial Net work (CSPNet) structure to propose a new CSP-Xception backbone network. The CSPNet network is capable of imp roving the learning ability of the network by using gradient split to build rich gradient co mbinations. Secondly, we design the Double Weighted Feature Fusion (DWFF) in the ASPP module. Weighted feature fusion by learning the importance of input features, and in this way, the segmentation accuracy of images with small size objects and complex background will be improved. The experimental results of training and testing on X-Ray security images show that the mPA value and mIoU value of CWF-DeepLabv3+ network reach to 90.85% and 81.64% respectively. Furthermore, co mpared with DeepLabv3+, Unet, and Bisenetv2 semantic segmentation network, the mean pixel segmentation accuracy of our proposed network is improved by at least 3.13%.