Early-stage fertilization detection in pigeon eggs during incubation using improved lightweight model

Early detection of fertilized and unfertilized eggs in the early stage of poultry egg hatching can improve eco-nomic efficiency by saving space, reducing costs, and obtaining unfertilized egg values earlier. As such, detecting early hatching information of breeding eggs is crucial in the poultry industry. This paper proposes an improved Mov3-CapNet lightweight model combined with machine vision for accurate, fast and non-destructive discrimination of pigeon breeder egg fertilization. The model is based on the MobileNetV3, with some SE (Squeeze-and-Excitation) modules replaced by DFC (Coupled Full Connected) modules, the classification layer replaced with CapProNet (Capsule Projection Network) and part of the original channel extension layer censored. The experiments use an industrial camera to collect high-definition images of breeder eggs from day 1 to day 4, which are then preprocessed with data augmentation methods. The improved Mov3-CapNet lightweight network is trained and detected, achieving 95.54 % accuracy on day 1, 98.85 % accuracy on day 2, 99.73 % accuracy on day 3, and 100 % accuracy on day 4. The average detection time of a single sample using CPU and GPU is 30 ms and 0.7822 ms, respectively. This study has practical application significance for accurate, fast and nonde-structive determination of pigeon breeder egg fertilization.