Semantic segmentation algorithm of underwater image based on improved DeepLab v3+

Abstract

Underwater images are easily affected by external factors such as light and water impurities, resulting in low segmentation accuracy of underwater images. Aiming to solve the problem, this paper proposes a semantic segmentation method of underwater images by introducing a new branch to improve the structure of the DeepLab v3+ model. It exploits the low-semantic features of the input image to improve the performance of the network. In the new branch, down-sampling is used to adjust the resolution of the input image, and the dimension of the feature map is adjusted by 1×1 convolution. The attention mechanism is used to focus on the channel of the feature map, and finally, the feature map of the existing branch is merged. In the experiments on the SUIM underwater dataset, the average intersection ratio of the model is 72.88%, and the average pixel accuracy is 84.03%. In the experiments on the PASCAL VOC dataset, the average intersection ratio of the model is 75.85%, and the average pixel accuracy is 84.5%. Compared with existing mainstream algorithms, the proposed algorithm achieves better results.