Automated counting and identification for low-yield hen cages in large-scale farming

We developed a simple yet effective, plug-and-play combination compatible with any model, featuring Mixed Local Channel Attention (MLCA), Light Efficient Detection (LED), and Weighted Intersection over Union (WIoU) modules. Poultry houses rely on manual checks to find low-yield hens, a subjective and labor-intensive task. We used an inspection robot to collect real-world data from 270 three-tier cages of 'Jing' hens. Integrating these modules enhanced the capture of small and overlapping targets like hen heads and eggs, boosting processing speed. We also custom-designed a Counting-in-Different-Cages (CDC) algorithm to solve counting difficulties caused by the tiered cages. The improved model increased hen detection precision by 4.7%, recall by 5.4%, and processing speed by 20.7 frames per second. In actual counting tasks, the algorithm achieved 93.1% accuracy for hens and 97.5% for eggs, with a minimal error of 0.26 hens and 0.06 eggs per cage. The entire system identified low-yield cages with 93.3% accuracy, proving it is highly capable of automated performance monitoring in commercial poultry farms.