Study on a Resnet18-Based Lightweight Recognition ALgorithm of Coal and Gangue

The pre-sorting of gangue in raw coal is still predominantly operated manually. The process is characterized by high labour intensity, low efficiency and possible safety risks. A robot designed for the sorting of the gangue in the raw coal is an effective way of completing this task without the need for manual labour. The key technology behind this is the intelligent identification of coal and gangue. The real-time performance has a significant impact on the efficacy of the gangue sorting robot. A multitude of models have been put forth for the recognition of coal and gangue based on machine learning, but remaining several issues, including slow recognition speed and large model size. To address those problem, this paper proposes a recognition algorithm with a backbone feature extraction of ResNet18 for coal and gangue during coal preparation, ResNet18-YOLO. It also studies the feature scale clipping and unstructured pruning of the model with the aim of further improving its recognition speed. Coal and gangue dataset were prepared based on the on-site gangue sorting situation. The investigation results demonstrate that the enhanced ResNet18-YOLO is capable of detecting coal and gangue, and exhibits superior real-time performance, thereby showing an optimum balance between accuracy and real-time detection of coal and gangue. Also commendable is its performance in detecting coal and gangue.