A high accuracy detection method for coal and gangue with S3DD-YOLOv8

Abstract

ABSTRACT In the process of coal sorting, the similarity in color and shape between coal and gangue makes the sorting process more difficult. This paper proposes a new method for detecting coal gangue based on an improved version of YOLOv8 (You Only Look Once). First, to enhance the recognition of different types, the input images are converted from Red-Green-Blue (RGB) channel to Hue-Saturation-Value (HSV) channel. Then, a Squeeze-and-Excitation (SE) attention module is added to the first layer to increase the robustness of the network. Next, a 3D dilated convolutional module is added after the SE attention module to improve feature detection. Furthermore, the SE attention mechanism module has been integrated with the existing C2f module. Finally, the last 2D convolution in the backbone is replaced with a deformable convolution (DCN). In this paper, we call new model S3DD-YOLOv8. The experiment on self-built coal and gangue data sets showed that the enhanced YOLOv8n has significant advantages in the detection. The Precision and Recall have been increased by 11.95% and 7.769%, and the mAP50 has been improved to 0.988. Our algorithm can effectively detect coal and gangue that are stacked together in various quantities and sizes under various lighting conditions, and demonstrates a certain level of robustness and practicality.