Abstract: YOLOv3 is a deep learning target detection algorithm with high detection accuracy and fast recognition speed.Using the improved YOLOv3 algorithm with depth-separable convolution to improve the accuracy and efficiency of coal-rock interface recognition, which is the key technical bottleneck of intelligent mining in coal mines.Aiming at the problem that the traditional target detection evaluation index cannot accurately evaluate the accuracy performance of the algorithm recognition caused by the continuous and penetrating characteristics of the coal-rock interface,the new coal-rock interface recognition accuracy evaluation index has been established.The cubic spline interpolation algorithm is applied to fit the points in the coal-rock interface identification prediction frame,and a close-to-real coal-rock interface curve is obtained.Related experimental results show that the training parameter scale of improved YOLOv3 is reduced by about 80%,and the test time is reduced by about 5%.Using new evaluation indicators,the accuracy rates in the x and y direction are respectively increased by 5.85% and 16.99% and the extracted coal-rock interface curve is controlled within 4.1% of the true value error.