Leveraging SOLOv2 model to detect heat stress of poultry in complex environments

Heat stress is one of the most important environmental stressors facing poultry production. The presence of heat stress will reduce the antioxidant capacity and immunity of poultry, thereby seriously affecting the health and performance of poultry. The paper proposes an improved FPN-DenseNet-SOLO model for poultry heat stress state detection. The model uses Efficient Channel Attention (ECA) and DropBlock regularization to optimize the DenseNet-169 network to enhance the extraction of poultry heat stress features and suppress the extraction of invalid background features. The model takes the SOLOv2 model as the main frame, and uses the optimized DenseNet-169 as the backbone network to integrate the Feature Pyramid Network to detect and segment instances on the semantic branch and mask branch. In the validation phase, the performance of FPN-DenseNet-SOLO was tested with a test set consisting of 12,740 images of poultry heat stress and normal state, and it was compared with commonly used object detection models (Mask R CNN, Faster RCNN and SOLOv2 model). The results showed that when the DenseNet-169 network lacked the ECA module and the DropBlock regularization module, the original model recognition accuracy was 0.884; when the ECA module was introduced, the model's recognition accuracy improved to 0.919. Not only that, the recall, AP0.5, AP0.75 and mean average precision of the FPN-DenseNet-SOLO model on the test set were all higher than other networks. The recall is 0.954, which is 15, 8.8, and 4.2% higher than the recall of Mask R CNN, Faster R CNN and SOLOv2, respectively. Therefore, the study can achieve accurate segmentation of poultry under normal and heat stress conditions, and provide technical support for the precise breeding of poultry.