Research on Image Semantic Segmentation Method Based on Combination of DeepLabV3+ and Structure Reparameterization

Aiming at the problems of incomplete segmentation of small objects when the sample is not balanced and low accuracy in the current commonly used semantic segmentation models, an improved DeepLabV3+ network is proposed in this paper. By using the structural reparameterization method, this paper replaces the normal convolution in the partial structure of the network with DBB module, which allows the network to have a simple structure while maintaining the high segmentation performance that comes with complex multi-branch network. In addition, the network is trained by using a hybrid loss function consisting of cross entropy and dice loss function to further improve the segmentation accuracy. The proposed algorithm is verified on the Cityscapes dataset and the obtained mIoU value is 78.84% and MPA is 89.49%, which are 3.69% and 3.94% higher than the original model, respectively. Experiments show that the improved segmentation algorithm proposed in this paper can effectively improve the segmentation accuracy and achieve better segmentation effectiveness.