Improved DeepLabV3+ image Semantic Segmentation Algorithm

Abstract

For the existing image semantic segmentation algorithms that have some difficulties, such as small target missing segmentation, false segmentation, and imprecise boundary segmentation, an optimized DeepLabV3+ image semantic segmentation algorithm is proposed. Firstly, the backbone network selects Xception_65 and Xception_71 networks. After comparison, the Xception_71 network is selected; Then, a parallel multi-branch module (PMB) is proposed, and the hole convolution with hole rates of 6, 12, and 18 in the Atrous Spatial Pyramid Pooling (ASPP) module is changed to the hole convolution with hole rate of 4, 8, 12 and 16, to obtain low-level feature map data. The result will be better and ensure the high accuracy of image segmentation; In the decoding layer, the direct up-sampling multiple of the DeepLabV3+ algorithm will lead to the loss of pixel information in the image. Therefore, using a bilateral up-sampling decoder to increase the continuity of pixel information in the image; Channel attention is added to emphasize important features so that the model has a stronger generalization ability. The focal loss function is selected to eliminate the difficulty of sample imbalance in the segmentation target. Finally, experiments confirm that the mIoU value of the proposed method on the data set PASCAL VOC2012 reaches 85.4%, and the segmentation effect is significantly enhanced.