Abstract: To enhance the intelligence and efficiency of manual coal mining, this paper proposes a top coal caving control method based on visual perception.First, by analyzing the differences between coal and gangue under various illumination conditions, the illumination level with the maximum difference is selected for image acquisition.Next, considering the actual underground production conditions, a de-dusting algorithm based on the dark channel is introduced.Then, a semantic segmentation model with dual tasks for gangue and foreground is designed to classify the pixels in the images into foreground and gangue.After obtaining the classification results, the gangue ratio in the foreground is calculated to determine the mixed gangue rate.By setting a threshold for the mixed gangue rate, intelligent control of the coal mining process is achieved.Finally, the trained segmentation model is deployed at the 81202 working face.The segmentation results are tested to verify that the model's accuracy meets the actual production requirements.With a mixed gangue rate threshold set at 20%,an ash content analyzer is used to test the coal quality after deploying the model for eight cuts, showing a mixed gangue rate of 17.6%.This demonstrates that the proposed control method meets the precision requirements for production control.