A lightweight DDC-YOLO and inspection-robot system for dead broiler detection in caged broiler houses

In caged broiler houses, dead-broiler detection typically relies on manual inspection, which is labor-intensive, inefficient, and difficult for high-level cages. To address this issue, this study developed an automatic dead-broiler detection system for caged farming environments to support automated identification and inspection-based monitoring of dead broilers under complex commercial production conditions. First, to mitigate complex lighting and image blur, Gamma transform and Gaussian blur were applied for data augmentation to simulate real breeding scenes and improve model robustness. Second, based on the YOLO11 framework, a lightweight improved model, DDC-YOLO (Dead Chicken YOLO), was developed for dead-broiler detection in caged farming environments; it uses a UIB lightweight backbone, the WIoU loss, the SimAM attention mechanism, and a DBFF feature-fusion module to enhance feature extraction and target detection while reducing model complexity, combined with TensorRT on edge devices to accelerate inference. Finally, this study established a human-review-assisted model updating workflow that integrates inspection-robot-based data acquisition and detection, cloud-based transfer learning, and iterative model updating, providing an engineering pathway for adapting the model to different broiler breeds and new production scenarios after deployment. The experimental results showed that the proposed DDC-YOLO achieved a precision of 96.8 %, a recall of 85.4 %, an F1 score of 90.8 %, and an [email protected] of 93.1 %. Compared with the original YOLO11, the model reduced parameters by 1.437 M and decreased size by 2.8 MB. The proposed model has been deployed in a real farm environment, where it supports real-time dead-broiler detection and result reporting. Furthermore, the human-review-assisted cloud-edge collaborative updating workflow provides an engineering pathway for transfer updating with newly collected field samples during subsequent deployment. Overall, the proposed method balanced detection accuracy, edge-deployment capability, and continuous optimization potential, and it can provide technical support for unmanned and reduced-staff broiler cage farming.