Research on broiler health status recognition method based on improved YOLOv5

Abstract

Caged broilers, reared extensively, necessitate timely health assessment to mitigate farming losses. Therefore, accurate detection of broiler health status holds a pivotal role in averting avian disease propagation. In this study, we propose a method for broiler health recognition using YOLOv5′s original structure, augmented with the lightweight mobile-EMO network model. Initially, we replace YOLOv5′s backbone convolutional, pooling, and connectivity layers with the EMO module, except the decoupling header. This substitution crafts an efficient model, amplifying overall performance. The EMO module introduces iRMB, a more effective feature extraction technique capturing diverse-scale feature information. Moreover, a continuous activation function supplants the conventional ReLU, enhancing feature expression. Further, EMO integrates 1×1 inflated convolution and windowed sub-attention, augmenting network performance. Additionally, we employ the SPPF module instead of SPP, replacing Detect with Decoupled Detect, constructing an advanced YOLOv5 network for broiler health state recognition. Through training and evaluation, a deep learning model is formulated, discerning caged broiler health. Experimental outcomes reveal that, even under intricate conditions, health detection attains a 0.959 recall rate, 0.956 accuracy, and 0.978 average accuracy. In comparison with basic YOLOv5, this method heightens average accuracy by 1.5%. Compared to YOLOv7 and YOLOv8, average accuracy increases by 3.1% and 1.8% correspondingly. The proposed approach fosters preliminary intelligent health detection, curbing manual intervention, lowering breeding costs, augmenting farming economics, and referencing health intelligent detection in broiler breeding.