Abstract: In low-illumination mine environments, conveyor belt foreign object detection algorithms suffer from insufficient extraction of global image features and an excessive number of model parameters. A method for detecting foreign objects on mine conveyor belts based on an improved version of YOLOv8s was proposed. YOLOv8s was improved using VMamba and MobileNetv4: MobileNetv4 was employed to enhance the backbone network by integrating the Universal Inverted Bottleneck (UIB) module. The efficient inverted residual structure reduced the overall number of model parameters, and a dynamic feature adaptation mechanism was used to strengthen feature robustness in small-object scenarios. The core feature extraction and fusion module C2f was improved by VMamba's Visual State Space (VSS) module, which efficiently captured global contextual information in images through a state space model and four-directional scanning mechanism, enhancing the model’s understanding of global image structure. A parameter-sharing lightweight detection head was designed, using Group Normalization (GN) as the basic convolutional normalization block to compensate for accuracy loss caused by model lightweighting. Experimental results showed that the improved YOLOv8s model achieved an mAP@0.5 of 0.921 and an mAP@0.5:0.95 of 0.601 on a self-built dataset, reduced the number of parameters by 27.7% compared to original YOLOv8s, outperformed mainstream object detection models such as YOLOv11s and YOLOv10s, and met the requirements for foreign object detection on mine conveyor belts.