Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Highlights

A real-time video acquisition platform is built in the ranch, so images/videos can be acquired automatically; based on this platform, our dataset used for training and testing is built.

ABSTRACT. Under the requirements of smart farming and precision management, accurately detecting and classifying cow body conditions is crucial for the large-scale intelligent breeding of cows. This study proposed a novel body condition scoring model for cows based on the YOLOv7 (you only look once version 7) framework. First, a vision transformer with Bi-Level Routing Attention (BiFormer bidirectional feature transformer) mechanism was added to the reparameterized path of the YOLOv7 backbone network to enhance the feature extraction capability. Furthermore, the convolution in YOLOv7 was replaced with a learnable efficient convolution operator (DSConv (distributed shift convolution)). Finally, the NWD (normalized Wasserstein distance) loss module was integrated into the CIoU (complete intersection over union) loss to reduce the regression error of the original model. The experimental results demonstrate that the improved YOLOv7-BDN model achieves an mAP@0.5 of 95.9%, and an F1 score of 90.4%. Compared with the YOLOv5m, YOLOv6m, YOLOv7, YOLOv8m, and YOLOv8l models, these indicators improved by 4.7%, 4.0%, 2.2%, 4.7%, and 2.2%, respectively. In this work, Gaussian noise with a noise kernel of 18 is added, and data augmentation techniques are adopted in the training dataset. The YOLOv7-BDN model proposed in this study can accurately identify the body condition of dairy cows in complex environments with stronger robustness and generalizability.