Image-based quantification of chicken carcass size using deep learning and image processing to support screening in poultry processing.

Miyazaki Prefecture is one of Japan's leading regions in chicken production, where strict poultry inspection is conducted at processing plants. Licensed veterinarians serving as certified poultry inspectors visually examine approximately 50,000 chicken carcasses per day, raising concerns about inspector burden and inspection reliability. This study developed an image-based method to support screening of nonconforming carcasses, focusing on emaciation and wooden breast (WB). Video footage of the processing line was geometrically corrected and both carcasses and their supporting shackles were detected using You Only Look Once version 8 (YOLOv8). The trained model achieved a validation mean average precision at 50% intersection over union (mAP50) of 0.995. The line speed was estimated to track carcasses across frames, and known shackle dimensions were used as a reference scale to calculate carcass area and an aspect ratio defined as a thickness index. The thickness index reflects torso fullness and complements carcass area as an independent feature. Visual inspection by a former poultry inspector indicated consistency between the calculated indices and visual impressions such as "emaciated" or "WB". Scatterplots of area versus thickness index showed only a weak positive correlation, suggesting that these indices capture distinct aspects of carcass conformation. Applying threshold lines on the plots allowed classification of some nonconforming carcasses, demonstrating the potential utility of these parameters as supplementary inspection indicators. This approach may reduce inspector burden, standardize inspection quality, and contribute to the development of future automated poultry inspection systems.