A Lightweight Deep Learning Approach for Real-Time Detection of Egg-Laying Behavior and Hatching Eggs in Breeding Ducks

Egg production rate and laying behavior are important indicators of reproductive performance in breeding ducks. Manual monitoring is time-consuming and labor-intensive, and litter, small eggs, variable postures, and occlusion complicate detection. This study proposed DPG-YOLO to detect egg-laying behavior and hatching eggs. Based on YOLO26, the DE-C3K2 module was reconstructed to enhance fine-grained feature extraction. PPA was incorporated to strengthen multi-scale semantic modeling, and C3Ghost was adopted to optimize neck feature fusion and reduce redundant computation. DPG-YOLO achieved precision, recall, and mAP50 values of 96.8%, 96.3%, and 97.4%, respectively, exceeding YOLO26 by 4.6, 4.8, and 5.0 percentage points. The model required 7.6 GFLOPs, occupied 7.1 MB, and processed 124.1 frames/s. Automatically estimated egg production rate and laying behavior frequency were significantly and positively correlated with manually calculated values (rs = 0.987 and 0.964, p < 0.05). Neither indicator differed significantly between the two methods (p > 0.05). Thus, DPG-YOLO combines detection accuracy, real-time performance, and statistical reliability to support automated monitoring of reproductive performance, precision feeding management, and digitalization of breeding duck production.