Abstract: In order to enhance the level of intelligence in coal mine excavation working faces and address the issues of high sensor dependency, poor environmental adaptability, and low utilization of image information in traditional process identification methods, a machine vision-based key process identification method for coal mine roadway excavation face is proposed. By analyzing the actual excavation process, the four typical key operations of “walking, cutting, temporary support, and permanent support” were identified. Combining key features such as target category, target motion state, and personnel number, a procedure discrimination strategy consisting of 11 key features was constructed. Based on this, the CSIS-YOLOv11 object detection model is proposed as the core visual recognition module of this method. This model incorporates three key improvements over YOLOv11n: First, it introduces the Contrast Limited Adaptive Histogram Equalization (CLAHE) image enhancement algorithm to improve the contrast and edge clarity of images in low-light, dusty, and foggy complex environments, making the image features more prominent and providing higher detectability for subsequent algorithms. Second, the SEAM attention mechanism module is introduced in the detection head, utilizing multi-scale channel modeling and feature fusion strategies to enhance the perception capability in multi-target occlusion scenarios, and to improve the model’s adaptability to local feature loss and inter-class interference, thereby increasing the robustness and recognition accuracy of target detection. Third, an IS-IoU loss function is designed by integrating the principles of Inner-IoU and Shape-IoU, introducing auxiliary bounding boxes to constrain the regression process while accounting for shape and aspect ratio variations, which improves localization accuracy, especially for elongated or scale-variant targets. Combining self-collected and annotated coal mine excavation image datasets for experiments, the results show that the improved CSIS-YOLOv11 model achieves an average precision (mAP) of 0.929, an increase of 3.6% compared to the original YOLOv11n model, while also featuring a lightweight structure (2.41×10⁶ parameters), low computational complexity (FLOPs is 5.5×10⁹), and high real-time performance (detection frame rate of 81.7 fps). Based on this, a excavation key process identification system based on the B/S architecture is constructed, integrating functional modules such as video capture, image enhancement, target detection, process logic inference, and visualization display, supporting real-time identification and duration statistics of key processes. The system carries out multi-process video recognition verification under laboratory conditions, and the recognition accuracies are 100% for unworked or other processes, 91.40% for walking, 99.36% for cutting, 98.39% for temporary support, 99.25% for permanent support, and the overall average recognition accuracy reaches 97.84%. The experimental results show that the key process recognition method can meet the requirements of intelligent detection of key processes in the face of roadway excavation in coal mines.