Automatic detection of brown hens in cage-free houses with deep learning methods

Computer vision technologies have been tested to monitor animals' behaviors and performance. High stocking density and small body size of chickens such as broiler and cage-free layers make effective automated monitoring quite challenging. Therefore, it is critical to improve the accuracy and robustness of laying hens clustering detection. In this study, we established a laying hens detection model YOLOv5-C3CBAM-BiFPN, and tested its performance in detecting birds on open litter. The model consists of 3 parts: 1) the basic YOLOv5 model for feature extraction and target detection of laying hens; 2) the convolution block attention module integrated with C3 module (C3CBAM) to improve the detection effect of targets and occluded targets; and 3) bidirectional feature pyramid network (BiFPN), which is used to enhance the transmission of feature information between different network layers and improve the accuracy of the algorithm. In order to better evaluate the effectiveness of the new model, a total of 720 images containing different numbers of laying hens were selected to construct complex datasets with different occlusion degrees and densities. In addition, this paper also compared the proposed model with a YOLOv5 model that combined other attention mechanisms. The test results show that the improved model YOLOv5-C3CBAM-BiFPN achieved a precision of 98.2%, a recall of 92.9%, a mAP (IoU = 0.5) of 96.7%, a classification rate 156.3 f/s (frames per second), and a F1 (F1 score) of 95.4%. In other words, the laying hen detection method based on deep learning proposed in the present study has excellent performance, can identify the target accurately and quickly, and can be applied to real-time detection of laying hens in real-world production environment.