Hybrid Patchcore-YOLOv8 Framework for Anomaly Detection in Chicken Egg Embryos

Abstract

Efficient and accurate assessment of egg fertility and embryo viability is crucial in the poultry industry. Manual candling is subjective and inefficient for detecting subtle developmental anomalies, particularly within the critical first week of incubation. This paper proposes an automated, non-destructive system using a hybrid deep learning framework, "YOLOv8+PatchCore," for anomaly detection in chicken embryos. The system leverages images captured via smartphone candling of eggs within a standard incubator. This two-stage approach first uses YOLOv8 as the "Eyes" to rapidly localize the egg (Region of Interest, or ROI) within the image. Subsequently, Patchcore acts as the "Brain," an unsupervised anomaly detector that analyses the provided ROI to identify subtle, unknown developmental deviations such as blood rings or arrested development. This framework, trained on a custom dataset, demonstrated exceptional performance. The integrated YOLOv8+PatchCore framework achieved an image-level anomaly detection AUROC of 0.96 and an overall accuracy of 89.10%, outperforming the baseline Patchcore-Only approach (AUROC of 0.92). This integrated system offers a practical, scalable, and highly accurate solution for automated quality control in commercial poultry production.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint