In the steel metallurgy industry, with the advancement of digitalization, networking, and intelligence, a large amount of image data from the sintering process remains underutilized. This paper proposes a sintering surface defect detection model, YOLO_LSK, based on a large selective kernel network (LSK), which utilizes computer vision technology for real-time detection of surface defects. The model uses an improved YOLOv7 for defect detection, incorporating the LSK feature extraction module to dynamically adjust the receptive field, allowing for the accurate identification of both moderately severe and severe cracks. With the iteration of training epochs, the recognition accuracy for these two types of cracks reached 100%, and the model's mAP achieved 63.17%. This study constructed a dataset of 1,000 real sintering process images, and through manual annotation and expert experience, established a qualitative and quantitative defect evaluation model to ensure the accuracy and reliability of the detection results. Experimental results show that the YOLO_LSK model excels in real-time crack length and direction identification, significantly improving the quality and efficiency of steel production.