Abstract
Feather sexing, based on wing feather growth rate, is a widely used method for chick sex identification. However, it still relies on manual sorting, necessitating automation. This study proposes an improved SAG-YOLO method for chick sex detection. Firstly, the model reduces both parameter size and computational complexity by replacing the original feature extraction with the StarNet lightweight Backbone. Next, the Additive Convolutional Gated Linear Unit (Additive CGLU) module, incorporated in the Neck section, enhances multi-scale feature interaction, improving detail capture while maintaining efficiency. Furthermore, the Group Normalization Head (GN Head) decreases parameters and computational overhead while boosting generalization and detection efficiency. Experimental results demonstrate that SAG-YOLO achieves a precision (P) of 90.5%, recall (R) of 90.7%, and mean average precision (mAP) of 97.0%, outperforming YOLO v10n by 1.3%, 2.6%, and 1.5%, respectively. Model parameters and floating-point operations are reduced by 0.8633 M and 2.0 GFLOPs, with a 0.2 ms faster GPU inference speed. In video stream detection, the model achieves 100% accuracy for female chicks and 96.25% accuracy for male chicks, demonstrating strong performance under motion blur and feature fuzziness. The improved model exhibits robust generalization, providing a practical solution for the intelligent sex sorting of day-old chicks.