Abstract: Manual interpretation of variable-density log (VDL) images is still the mainstream approach for assessing the bond quality of the cementing second interface (the cement sheath–formation interface), but it is time-consuming, strongly subjective, and often inconsistent. To improve evaluation accuracy and efficiency, we proposed Multi-Scale Feature Hybrid Channel–Spatial Attention Network (MSF-HCSA Net), a convolutional neural network that integrates multi-scale feature extraction with a hybrid channel–spatial attention module to automatically evaluate second-interface bond quality from VDL images. This model was trained and validated based on the data from three wells in the Shunbei Oilfield. The evaluation accuracy of the second interface reached 95.8%. In the case where the sample was unbalanced, and the proportion of small samples with “poor bond quality” was low, the general convolutional model SLaK had deficiencies in recognizing such samples. In contrast, MSF-HCSA Net utilized channel-spatial hybrid attention and multi-scale feature fusion to increase the recognition accuracy of the “poor bond quality” category in small samples by 10%. To a certain extent, this alleviated the performance degradation caused by the imbalance between classes. The research results show that the proposed MSF-HCSA Net can achieve rapid, objective, and efficient automatic evaluation of the quality of the cementing second interface, providing reliable technical support for on-site cementing quality monitoring and follow-up optimization.