HighlightsRapid and accurate identification and detection of mortality in high-level caged broilers.Apply the state-of-the-art single-stage object detection model Yolov5 as the experimental infrastructure.Modify and improve the network structure, reduce the calculation amount of model parameters, and make the model easier to deploy.The data of dead broilers in different regions can also be quickly and accurately identified and detected, with excellent generalization.Abstract. In an unmanned broiler base, the automatic identification of the death state of broilers in a complex cage environment can effectively improve picking efficiency and reduce its negative impact. An improved Yolov5 object detection method is proposed in this study to rapidly and accurately detect the presence of dead broilers in caged structures. The Yolov5 model within the CSPDarknet53 framework is constructed, and the Yolov5 model is experimentally pruned to reduce the number of convolution kernels. The Ghost structure and the Involution structure are combined with the Yolov5 backbone network to effectively reduce the number of parameters and computations with a negligible loss of accuracy, making it easier to deploy applications. The detection accuracy of the model can reach 99.2%. The number of parameters is reduced from 7.05M to 0.31M, and the amount of calculation is reduced from 16.3 to 0.3. It can be seen that the improved Yolov5 network is practical and suitable for high-precision real-time detection and recognition of the death state of broilers in a caged structure. Keywords: Keywords.,Computer vision, Dead chicken detection, Machine vision, Yolov5.
Read full abstract