Abstract: Surface nodulation is a major quality defect in electrolytic copper cathode products. In production practices, the problems that occur during the electrolytic process are often diagnosed according to the analysis of different types of nodules on the cathode copper plates. The traditional manual observation method for determining nodule types on copper cathode plates has the disadvantages of low accuracy, time lag, etc. An improved DeeplabV3+ semantic segmentation model was proposed, which can be deployed on-site to achieve real-time identification of nodule types on the surfaces of copper cathode plates. MobileNetV2 was the backbone network to achieve lightweighting, with a model size of 11.15% of its original size. A spatial and channel attention mechanism was introduced to capture multi-scale information to improve the accuracy of nodule edge region segmentation, resulting in a 1.06% increase in the accuracy of defect category classification. The experimental results showed the algorithm’s excellent segmentation and classification effects on electrolytic copper cathode plates’ point-like, clustered and edge nodule defects. The segmentation accuracy on the test set reached 91.58%, which could meet the actual production demands and provide a practical reference for further intelligent control of surface quality online detection of cathode copper plates in electrolytic copper production.