Abstract
A novel approach, designated YOLOv8-POS, is introduced to address the issue of false detections in coal-rock image recognition tasks, frequently caused by factors such as image defocus, dim lighting, and worker occlusion, and to further enhance the model's accuracy and reduce its complexity. The methodology introduces a C2f-PConv module, which ingeniously combines the strengths of C2f and partial convolution (PConv) to selectively process channels. This reduces unnecessary computational overhead while preserving the integrity of critical feature information, thus significantly cutting down on the model's parameters and computational demands. Additionally, an Overlapping Spatial Reduction Attention module is incorporated into the model's architecture to optimize the fusion of spatial features, substantially improving the handling of complex scenarios. The adoption of a slim-neck design further streamlines the computational and storage requirements, leveraging meticulously engineered lightweight modules to enhance the model's practical applicability. Empirical results demonstrate that YOLOv8-POS markedly improves performance on coal-rock image datasets, achieving an AP(50) of 77.1% and an AP(50:95) of 63.6%, while concurrently reducing the model's parameters to 2.60 M and the floating point operations (FLOPS) to 6.4 G. Comparative evaluations with other prominent algorithms confirm the superior performance of this refined approach, solidifying its advantage in practical deployments.