A convolutional neural network boundary extraction model integrating multiple attributes of geological and seismic profiles is constructed, which introduces multi-scale convolution, jump connection with curvature and acoustic impedance gradient constraints based on the improved U-Net structure, in order to enhance the continuity of the fault boundary and the identification of weak reflection zones. The model adopts a three-channel input of 256 × 256 pixels and realizes efficient computation by 1 × 1 convolution compression channel in the feature fusion stage. The experimental results show that the precision rate of 92.15%, recall rate of 89.48%, and IoU of 0.821 are improved by 4.21%, 3.81%, and 0.039, respectively, compared with that of U-Net; the IoU still remains at 0.681 under 0 dB SNR, higher than that of U-Net's 0.605, which verifies the model's robustness and superiority in a multi-noise environment.