Research on coal-pencil image segmentation method based on improved YOLOv8n-Seg

To address the challenges of complex detection scenarios in coal-pencil Audibert-Arnu dilatometer images, such as low edge segmentation precision and large model size, this paper proposes a lightweight instance segmentation model named DSW-YOLOv8n-seg. The approach replaces standard downsampling convolutions with depthwise separable convolutions (DSConv) to lower computational cost. Additionally, a SimAM-based attention module is integrated to help the network better emphasize relevant spatial cues. For refining bounding box predictions, a Wise-IoU (WIoU) strategy is introduced to adjust gradient contributions adaptively. On a custom-built dataset of coal-pencil thermal expansion, the proposed method achieves an mAP of 87.6%, with a model size of 0.87M parameters and 192.6 GFLOPs, demonstrating improved edge segmentation under lightweight constraints.