A dual-modal vision system for non-invasive real-time monitoring of broiler diarrhea under low-light conditions.

The prevention of broiler diseases largely depends on the accurate identification of typical characteristics of broilers. Diarrhea, as a typical indicator of broiler health, is particularly important to identify accurately. In this paper, through field investigations and communications with professional veterinarians, it was determined that the presence of fecal crust adhering to the cloacal region of broilers can be used as a marker for broiler diarrhea. Consequently, a dual-modal broiler diarrhea detection network based on bimodal data fusion and attention mechanism (DBS-YOLO) is proposed. Firstly, considering the dim lighting conditions in most poultry farms, a bimodal broiler diarrhea dataset based on infrared and visible light images was established in this study. Secondly, to extract features from bimodal data, a Dual-backbone Feature Extraction Network (DFE-Net) was proposed. Subsequently, to filter feature information from different modalities, a Bimodal Adaptive Fusion Module (BAFM) was introduced. Moreover, this study innovatively proposed an attention-based feature selection module (C3-S), which, in combination with the Convolutional Block Attention Module (CBAM) attention mechanism, further enhanced the model's ability to fuse features of different scales. Finally, DBS-YOLO was compared with mainstream object detection algorithms. The experimental results showed that in terms of detection performance, DBS-YOLO achieved an mAP@0.5 of 97.2%, an mAP@0.95 of 57.3%, and an FPS of 96.46. This study provides new ideas for the prevention and detection of animal diseases in complex environments and lays the foundation for the research of intelligent poultry breeding equipment.