Abstract: The spiral concentrator is a widely used flowing-film gravity concentration device in metal and non-metal mineral processing. However, its concentrate grade adjustment heavily relies on manual observation and operation of the concentrate zone. To address the lag and errors in concentrate regulation caused by variations in human experience, which lead to fluctuations in processing indicators and loss of useful minerals, there is an urgent need to develop a rapid and accurate image recognition technology for identifying the concentrate zone. This paper proposes the Res-AFFUNet algorithm based on the UNet architecture to precisely segment ambiguous mineral zones. First, a ResNet50-CBAM module replaces the original UNet encoding feature extraction network, enhancing the model’s feature extraction capability. Second, the introduction of the dynamic upsampling DySample module improves segmentation edge accuracy. Subsequently, the attention-based feature fusion AFF module facilitates information integration between shallow and deep features. Finally, a Focal-Dice Loss hybrid loss function is constructed to effectively address the sample imbalance issue. Experimental results demonstrate that the proposed Res-AFFUNet algorithm achieves 99.07% precision, 96.32% mIOU, 98.69% mPA, and 98.70% F1 score in identifying spiral concentrator zone images. Compared to the UNet baseline, these metrics improve by 6.22%, 9.15%, 4.87%, and 5.54%, respectively. Furthermore, when compared against mainstream algorithms such as PSPNet, HRNet, and Deeplab V3+, our improved algorithm demonstrates significant advantages in mIOU, mPA, and F1 scores. In experiments evaluating mineral zone segmentation performance, the improved algorithm achieves more precise segmentation than other mainstream algorithms, effectively resolving the challenge of accurately identifying the blurred boundaries of spiral concentrator concentrates and enabling effective recognition of the concentrate zone. This study provides an efficient and precise solution for concentrate zone segmentation in spiral concentrators, laying the foundation for further development of concentrate grade control and adaptive extraction technologies for spiral concentrators.