Abstract: Accurately locating the salt structure buried underground is of great significance for efficient exploration and exploitation of oil and gas resources. Traditional deep learning methods still have shortcomings in the accuracy of salt boundary recognition and the ability to restore details. To improve the performance of salt body recognition, a VAC-U-Net network model based on an improved U-Net architecture is proposed. This model uses the first 13 layers of the VGG16 network as encoders to extract image features; By incorporating the ASPP module with residual connection mechanism, the ability to capture multi-scale contextual information is enhanced; Then, the CGAFusion feature fusion module, which combines three-level attention mechanisms of channel, space, and pixel, is introduced to effectively integrate multi-layer information of key regions and boundary features, and enhance the interaction ability of high-level and low-level semantic information; Finally, salt segmentation is achieved through a multi-level upsampling and decoding structure. Validated on the TGS salt body dataset, with an intersection over union of 85.49%, pixel accuracy of 96.21%, and F1-score of 91.84%. Compared with the original model, the model showed significant improvements in pixel accuracy and boundary restoration, demonstrating better robustness and generalization ability, providing effective technical support for underground salt body recognition.