Research on salt body recognition based on DeepLab V3 combined with auxiliary classifier and attention module

Salt bodies are a class of salt-containing minerals, its density is lower than that of the surrounding rocks. Relative to other rock formations, it behaves differently in acoustic reflection. According to this property, images of salt-containing minerals can be obtained by using technical means. In the Salt Label tasks, the semantic segmentation model of deep learning can be used to achieve pixel-wise classification of images. There are fewer existing methods of salt body recognition, therefore, we proposed a method adding an auxiliary classifier and attention mechanism to the DeepLab V3 baseline, which is an effective salt recognition network. Experiments show that auxiliary classifier can speed up convergence and increase accuracy on large networks. The attention mechanism increases the accuracy rate while steadily accelerating the convergence rate. When the backbone network is the MobileNet, adding the above two structures can increase the PA of the original network by 2% and mIou by 4.1%. When the backbone network is ResNet50, the PA can be increased by 4.6%, and mIou can be increased by 11%.