Abstract: To address the problems of unstable mineral processing accuracy and resource waste caused by manual adjustment of the concentrate splitter position in shaking table-based concentrate collection operations at mineral processing plants, an intelligent control system for shaking table-based concentrate collection relying on machine vision and deep learning was designed. The system adopted a three-level architecture consisting of "preparation layer, control layer, and application layer", establishing a closed-loop automated process from image acquisition and mineral belt recognition to concentrate splitter control. At the algorithmic level, the OreBound-YOLO mineral belt boundary point recognition algorithm was proposed. Based on the YOLOv7 network, this algorithm reconstructed the backbone network with the C3k2 convolution module to improve feature extraction efficiency, introduced the feature pyramid network + path aggregation network （FPN + PAN） structure to achieve bidirectional multi-scale feature fusion, embedded the C2PSA attention mechanism to enhance perception of weak boundary regions, and adopted the EIoU loss function to optimize bounding box regression accuracy. Industrial test results show that the system operates stably, effectively improving concentrate grade （the WO₃ grade of the concentrate increases by 2.4–2.9 times, and the Sn grade increases by 1.4–2.1 times）, reducing tailings grade, and significantly improving ore resource utilization, while greatly reducing reliance on manual operations and lowering enterprise operating costs. The research results provide a reliable technical path and system solution for realizing intelligent control of the shaking table-based concentrate collection process and have practical significance for promoting the automation upgrade of the mineral processing process and efficient resource utilization.