To accurately detect foreign objects on coal mine belt conveyors under complex working conditions, a coal mine conveyor belt foreign object segmentation model based on improved DeepLabv3+ was constructed. Aiming at the difficulties in foreign object detection caused by interference factors such as high dust, uneven illumination, and mechanical vibration in coal mines, as well as practical requirements including the coexistence of multi-scale foreign objects and limited computing power of edge devices, the following improvements were made: The MobileNetv3 lightweight backbone network was introduced, and depthwise separable convolution was used to compress the computational load to 1/9 of that of traditional convolution. Meanwhile, the SE attention module was embedded to enhance high-frequency features such as edges and textures of foreign objects and suppress low-frequency channels corresponding to dust noise. The DASPP module was adopted to replace the traditional ASPP, and cross-layer feature dense interaction was realized by concatenating atrous convolution layers with different dilation rates, thereby improving the detection for multi-scale foreign objects. The ECANet channel attention mechanism was integrated, which enhanced feature expression ability through dimension-reduction-free global pooling and dynamic 1D convolution, further optimizing the distribution of feature weights. Experimental results show that the improved model achieved an average mean intersection over union of 87.1% and an F1-score of 86.7% on the CUMT-BelT dataset, with only 9.8 M parameters, 5.1 GFLOPs, and an inference speed of 38.6 fps. Compared with the original DeepLabv3+ model, the improved model increased accuracy by 4.6% and reduced computational load by 63.1%. In comparison with mainstream models such as PSPNet and U-Net, the improved model exhibits superior performance in key indicators including the missed detection rate of small-scale foreign objects, noise robustness, and adaptability to edge devices. This model provides a new approach to solving the problem of easy confusion between foreign objects and background features under complex working conditions. By simultaneously achieving high segmentation accuracy and low computational complexity, the proposed method supports real-time monitoring and contributes to advancing automation and intelligent development in the coal industry.