关键词: 语义分割, SimAM, 三维深度可分离卷积, 断层识别

Abstract:  At present, deep learning-based fault recognition methods are emerging in an endless stream, focusing on the U-shaped network. However, the U-shaped network uses a large number of conventional convolutions, ignoring the problems of feature redundancy and over-fitting while improving the effect of feature extraction, resulting in high network complexity. In order to ensure the recognition effect and reduce the problem of feature redundancy, this paper proposes a lightweight fusion semantic segmentation network (LF-SeNet) for three-dimensional fault recognition. Compared with the traditional fault recognition network, LF-SeNet combines the idea of skip connection and feature fusion. The feature fusion module includes three-dimensional separable convolution, simple attention mechanism (SimAM), Dropout layer and finite matrix product operation, forming a lightweight feature fusion module. The effect of feature extraction is effectively guaranteed. In order to effectively reduce the complexity of the network, this paper combines the dilated convolution and the lightweight feature fusion module. On the one hand, it reduces the computational complexity of the network, and on the other hand, it reduces the feature redundancy problem caused by conventional convolution. In addition, this paper uses the Dropout layer and data enhancement method to improve the generalization ability of the network and alleviate the problem of overfitting. The method is tested on the FaultSeg3D dataset. The results show that the parameter amount of LF-SeNet is 2.56M, the number of floating-point operations is reduced by 95.59G, and the intersection-over-union ratio is increased by 2% compared with the traditional U-shaped network. Finally, this paper uses three-dimensional data synthesis technology to visualize the fault identification map. The experimental results show that the faults identified by LF-SeNet are continuous and clear, indicating that the network has better generalization ability, which proves the effectiveness of LF-SeNet in fault identification. 

Key words: semantic segmentation, SimAM, three-dimensional depth separable convolution, fault identification