Algorithm for Coal Mine Rock Foreign Object Detection Based on Enhanced YOLOv8

Abstract

In addressing the challenge of suboptimal detection precision in coal mine rock foreign object detection due to complex environments and variable object scales, this paper proposes a coal mine rock foreign object detection algorithm based on YOLOv8. Initially, the incorporation of the BiFormer attention mechanism is advocated to refine the backbone network, augmenting the model's attention towards pivotal information regions, consequently enhancing localization and feature extraction capabilities. Secondly, a lightweight Content-Aware Recurrent Affine Feature Extraction (CARAFE) operator is utilized within the neck architecture to effectively capture and preserve intricate features at lower hierarchical levels. Finally, Wise-IoU v3 is adopted as the bounding box regression loss for the proposed algorithm, coupled with a prudent gradient allocation approach, thereby enhancing the model's localization capabilities. Empirical findings illustrate that compared to baseline algorithms, the proposed algorithm has fewer parameters, with an average mAP improvement of 2.5%, and a detection speed increase of 2fps/s.