An Improved YOLOv5 for High-Resolution Ore Image Detection

During ore dressing production, the size of the ore significantly affects the grinding and classification efficiency and is a crucial control parameter. In the past, most ore image detection algorithms focused on the detection tasks of high-resolution ore image. At the same time, due to the sparse data set for industrial production applications, the generalization ability and robustness of the model are low, and it has poor performance on tiny ore target detection tasks. To remedy these shortcomings, this paper proposes the ODA-YOLOv5 network model for ore image detection. In this paper, we adopt a multi-scale training strategy to address the difficulty of detecting tiny ores in high-resolution images. A self-built dataset is constructed by fusing multi-scale ore images. The input image size is dynamically adjusted during training to enhance the robustness of the detection model for ore of different scales. In order to improve the performance of the detection algorithm, we introduce the C3-ECA attention module, which fully takes into account the local cross-letter road interactions and considerably reduces the model complexity. Our algorithm achieves approximately 97.9% mAP and approximately 62.11 FPS on the self-built dataset. Our algorithm sufficiently demonstrates the superiority of our detection model algorithm for industrial ore detection.