Improving Crack Detection Accuracy in Candled Chicken Egg Images Using Modified YOLOv8 with DCNv2

Abstract

Chicken eggs are essential poultry products in the food industry, and their quality determines consumer trust and selling value. Egg quality assessment is still generally performed manually using the time-consuming, subjective, and inconsistent candling method, making it inefficient on an industrial scale. In this study, an automated system based on computer vision and deep learning was developed to detect and classify eggs into two categories, cracked and intact. The YOLOv8 model serves as a base, and its performance is further enhanced through the integration of the deformable convolutional network v2 (DCNv2), which improves feature extraction, particularly for fine cracks that are difficult to detect. The egg images were collected using a Logitech camera (1080 HD, 30-60 fps) positioned 15 cm above a conveyor belt at three speeds (0.03 m/s, 0.04 m/s, and 0.05 m/s) to simulate industrial conditions. The YOLOv8-DCNv2 model shows better performance than the basic YOLOv8 model. Improvements are seen in several evaluation metrics, namely precision increased from 0.980 to 0.992, recall from 0.948 to 0.999, and mAP50 from 0.994 to 0.995. These results indicate that the addition of DCNv2 is able to improve the model's accuracy in detecting objects, especially fine cracks. The integration of DCNv2 has been proven to provide more efficient, consistent, and reliable detection performance, making it a potentially optimal solution for egg quality inspection systems in the modern poultry industry.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint