Development of an image analysis system at the slaughterhouse to assess laying hens’ welfare on farm

In the poultry sector, artificial intelligence could be a valuable means to ensure better farm welfare conditions and meet consumers’ demands, which are increasingly directed towards the choice of ‘welfare-friendly’ products. Much of the research in the field is focused on farm welfare where systems have been developed to detect behaviours, health or sounds automatically. However, some evaluations conducted directly on animals at the slaughterhouse may be iceberg indicators of poor on farm welfare conditions. The slaughterhouse represents a crucial and strategic point where evaluations can be carried out efficiently and quickly as it receives many animals daily, processed under standardised conditions. Last but not least, technological innovation and automation could support and facilitate veterinary inspection in monitoring such large numbers. In farmed laying hens, one of the main animal welfare consequence is bone lesions, including keel bone fractures, which, in addition to causing pain and suffering in the animals, can affect production performance and consequent economic damage. The causes for keel bone lesions are many and complex. Despite commonly linked to high egg production, husbandry systems, genetic lines and management practices also play fundamental roles in the onset and frequency of this problem. To gain an understanding of the welfare conditions on the farm and hypothesise corrective actions in the situations of most significant risk, the present research project focused on the development of an image analysis system that allows the automatic recognition and classification of keel bone lesions in laying hens. To this purpose, laying hens of brown genotypes in the slaughter line of a commercial plant in north-eastern Italy were filmed using a smartphone. 600 frames containing approximately five hens each were extracted from the videos obtained. On video, keel bone detection and lesion assessment were carried out by three different annotators using a three-score scale (0;1;unclassifiable). A total of 1443 images were obtained by applying the preprocessing and data augmentation steps. Finally, a YOLO11 model was trained. A mAP@50 value of 0.74 was obtained for the keel bone detection phase and an F1-score of 0.71 for the classification phase. The results obtained are preliminary and subject to improvement. The following steps will involve improving of the image dataset.