Aiming at the problem that mutton multipartite needs to be further classified and detected in the conveyor belt scene, a real-time classification and detection method for mutton splits based on a single-stage object detection algorithm was proposed. In the sheep slaughter workshop environment, multiple types and multiple mutton split images were collected. After image augmentation and normalization, a mutton multipartite image data set was established, including 7200 training sets, 1400 test sets, and 400 verification sets. Using the single-stage object detection algorithm YOLO v3 to introduce transfer learning to train the mutton multipartite image data set and obtain the optimal model. Based on the optimal model, the category and position of each mutton split in the image were returned, so as to realize the classification and detection of mutton multipartite. The average accuracy mAP and the average detection time of a single image were selected as the accuracy and speed indicators for judging the detection effect of the model. Then, the detection speed was optimized by replacing the feature extraction network of the mutton multipartite recognition model. In addition, an additional illumination data set containing two brightness levels of “bright” and “dark” and an additional occlusion data set representing the occlusion situation of mutton were set to verify the generalization ability and anti-interference ability of the optimized model，and the robustness of the optimized model was tested through the neck and abdominal rib with obvious multi-scale features. Finally, four commonly used object detection algorithms: Mask R-CNN, Faster R-CNN, Cascade R-CNN, and SSD were introduced to conduct comparative experiments on different data sets. On this basis, the feature extraction network was further replaced with MobileNet V1, ResNet34 and ResNet50 to verify the optimized model’s comprehensive testing capabilities. The test results showed that the detection speed of the optimized model was 48.53% higher than that of the original model. At the same time, it had strong generalization ability and anti-interference ability for multi-part recognition of mutton under complex environment of light and shading, and it had good robustness to the mutton multi-part detection with multi-scale features. It was optimized for the verification set of mutton multipartite image. The mAP value of the optimization model reached 88.05%, and the processing time of a single image was 64.7ms, the comprehensive detection ability was better than that of other algorithms, which indicating that this method had high detection accuracy and good real-time performance, and can meet actual production needs.