A lightweight object detection algorithm for coal and gangue with laser speckle imaging

Laser speckle imaging, known for its ability to capture object surface features with simple setup and reduced sensitivity to ambient light, is of significant research interest for coal and gangue recognition. However, the complex surface structures of minerals in practical settings pose challenges in improving recognition accuracy over large sample sizes and extended periods using manual feature design. To address this issue, we propose a coal and gangue recognition method that integrates laser speckle imaging with deep learning. Based on objective speckle imaging theory, we designed a coal and gangue laser speckle image acquisition system and curated a dataset encompassing diverse lighting conditions for speckle imaging. We developed a lightweight YOLOv5s model to extract rich surface information from mineral laser speckle images, achieving high-precision coal and gangue detection while reducing computational demands. Experimental results demonstrate significant improvements in model size, training effectiveness, feature extraction, and recognition accuracy by lightening the YOLOv5s model. Furthermore, our method exhibits improved accuracy and stability in coal and gangue recognition under varying lighting conditions compared to manual feature design approach. Additionally, our model strikes a balance between complexity and accuracy, offering practical advantages over existing models for industrial applications. These findings provide valuable technical support for the future realization of intelligent coal and gangue recognition.