RTDETR-Refa: a real-time detection method for multi-breed classiffcation of cattle 

Abstract

In the farming industry in order to cope with the problems such as the complex environment of the pasture and dense targets, etc., which lead to increased difficulty in recognition, so as to quickly classify and automatically identify cattle breeds and improve the accuracy. In this paper, an RTDETR-Refa (Rep- Conv Efficient Faster Attention) model is proposed for classification and identification of cattle breeds. Firstly, some novel improvements have been made to ResNet18. Replacing the 1×1 convolution in Faster-Block with the reparameterised RepConv, and adding the Efficient Multiscale Attention Module (EMA) in front of the global average pooling layer in Faster-Block to enhance the transformation and classification of features, Finally, the 4-layer BasicBlock after the 3 convolutional layers in the resnet18-backbone is replaced by the improved Faster-Block.Finally, the results of the training tests of the RTDETR-Refa model are compared with other classical models: the YOLO family of models: the YOLOv5m, YOLOv6, YOLOv8m, YOLOv9 and the state-of-the-art (SOTA) CNN backbone networks EfficientViT, FasterNet, UniRepLKNet, TransNeXt verifying their superiority. The average accuracy of the RTDETR-Refa model on the cattle classification training set is 91.6%, which is 0.8% higher than that of ResNet18 and 0.9-5.2% higher than that of other classical models. The experimental results show that the RTDETR-Refa model proposed in this paper is capable of identifying and classifying cattle of different breeds while ensuring similar detection speeds, demonstrating the feasibility of convolutional neural networks in breed identification and classification.