With the aim to solve the issues with poor accuracy of separation and delayed parameters adjustment in coal and gangue separation process, this study presents a model for coal and gangue separation that combines the DeepLabv3+ model and intelligent decision making. The model builds upon the DeepLabv3+ framework and incorporates the MobileNetv4 network to decrease parameter count and computational complexity while enhancing inference efficiency. Meanwhile, the Bottleneck Attention Module is introduced to strengthen boundary features between coal and gangue and achieve pixel-level extraction within coal and gangue regions. Moreover, the model integrates YOLOv7 to enhance multi-scale representations of coal and gangue and employs image coordinate mapping technology to enable rapid target localization and separation decision-making. Finally, this study uses a coordinated execution mechanism of a robotic arm and air blowing to achieve intelligent separation. Experimental results demonstrate that the model converges rapidly during training. Accuracy in recognition is 98.85%, while the average precision is stable at 98.64%. The model contains 30.82 M parameters and takes 17.27 ms to respond, both of which outperform the comparison models. Under dry conditions, the pixel accuracy of the three types of coal is 93.05% -94.87%, and the accuracy under high humidity is 90.74% -93.42%. Under the worst high-humidity brown coal conditions, the mAP@0.5 still reaches 91.28%, verifying the adaptability of the model. In the normal operation of the machine, coal precision is 96.83%. Even under high gangue content, coal precision is 93.76%. The above results indicate that this method has excellent coal rock recognition accuracy and sorting ability, and achieves a good balance between model complexity and accuracy, providing a feasible technical path for intelligent sorting optimization in coal preparation plants.
