DCDNet: A deep neural network for dead chicken detection in layer farms

Detecting deceased layers is vital in farm inspections. Manual inspections are inefficient and pose bio-security risks. Deep learning excels on large datasets, yet accurate dead chicken detection is challenging due to data scarcity, imbalance, visual similarity, and irregular morphology. To achieve desirable performance in distinguishing dead and normal chickens, a novel deep neural network, DCDNet, was proposed in this study. The pipeline consisted of the following three modules: Poisson fusion-based data augmentation (PFDA) module seamlessly integrated the area of the deceased layer into the new background, generating a more realistic image that alleviates sample scarcity; the designed DCDNet was utilized to accurately identify dead and normal layers by extracting and fusing features more efficiently, thus better suiting their irregular body shapes; the non-monotonic dynamic focusing (NDF) sliding weight loss function was proposed to enhance the contribution of difficult samples in model training flexibly, reducing bias caused by unbalanced data. Extensive experiments have been conducted on our dead-chicken dataset constructed on a commercial farm. The results revealed that the proposed method achieved a mean average precision (mAP) of 97.5%, outperforming the state-of-the-art methods reported thus far. Moreover, the average precision (AP) difference between dead and normal chickens is only 0.1%. The proposed dead chicken detection approach, based on DCDNet, was effective in dealing with sample scarcity and dataset imbalance. This may provide some reference for other researchers on other similar tasks.