Abstract: To achieve real-time detection of deboning locations during intelligent chicken feet deboning, this study addresses challenges such as poor segmentation and detection due to the curved morphology of deboning areas and the similarity between edge features and the background, as well as the high computational load and parameter count of existing models that hinder real-time detection on edge computing devices with limited capabilities. We proposed an improved lightweight segmentation and detection model based on YOLOv11n-seg, SCFL-YOLO. First, a fused lightweight StarNet architecture was adopted to improve the efficiency. Second, the DynamicTanh activation function was introduced into C2PSA to enhance feature representation. Third, in the neck network, the original convolution in C3K2 was replaced by a new fused convolution (FPConv) to construct the FPC3K2 module, thereby reducing model complexity while strengthening multi-scale edge feature extraction. Finally, a novel lightweight shared convolutional prediction head, LSCPLQS, was designed to enable multi-scale segmentation and detection of different deboning regions of chicken feet. Experimental results showed that SCFL-YOLO achieved a detection accuracy of 97.5%, a segmentation accuracy of 94.8%, and an inference speed of 180.8 fps. Compared with the baseline model, the parameter count, computational cost, and model memory footprint of SCFL-YOLO were reduced by 40.3%, 28.4%, and 39.7%, respectively, enabling effective segmentation and recognition of the claw pad and tarsometatarsal bone. Overall, SCFL-YOLO reduces computational complexity while maintaining high segmentation and detection performance, which can simplify the chicken feet deboning process and provide robust visual support for intelligent deboning equipment.

Key words: chicken feet deboning; detection and segmentation; lightweight; automation; deep learning