Abstract: To address the challenges of significant background interference, low segmentation accuracy, and high computational complexity in temperature anomaly detection of underground coal mine rollers, this paper proposes an improved DeepLabV3+ semantic segmentation model. First, the model design adopts the lightweight MobileNetV2 as the backbone network to reduce model complexity and meet the accuracy and deployment requirements of underground edge devices. Second, a Dual-Channel Attention Mechanism (DCAM) is integrated at the output of the original Atrous Spatial Pyramid Pooling (ASPP) module in the DeepLabV3+ network. Through dual-branch channel-spatial attention fusion of multi-scale thermal features, it suppresses complex background interference. Finally, a Temperature-Aware Weighted Cross-Entropy (TAWCE) loss function is designed, which assigns higher weights to high-temperature pixels, enabling the model to learn finer edge features and thereby improve the segmentation accuracy of abnormal region edges. Experimental results on a self-built laboratory infrared image dataset of rollers show that the improved algorithm achieves a mean Intersection over Union (mIoU) of 85.12%, which is 2.4% higher than the original DeepLabV3+, and the Mean Pixel Accuracy (MPA) for temperature anomaly regions increases to 87.92%. Additionally, the model parameters are reduced by 67.06%. This approach effectively enables rapid and precise segmentation of roller temperature anomaly regions, providing reliable technical support for intelligent operation and maintenance of coal mine belt conveyors.