Summary
YOLOv4 and YOLOv4-tiny were able to identify and classify DD lesions on images from a commercial dairy farm with high accuracy. Additionally, the YOLOv4-tiny were able to detect DD lesions in a milking parlor with high prediction speed as well. This result is a small step in applying computer vision algorithms to veterinary medicine and implementing real-time detection to dairy farms. The proposed computer vision tool can be used for early detection and prompt treatment of DD in dairy and beef cattle. The trained model can be applied to different cattle breeds and various locations. Additionally, such methods can be extended to other problem settings including classifying additional diseases in other animal species.
Overall, all models were able to detect DD lesions in dairy cattle on an image. All R-CNN and YOLO models improved performance over the YOLOv2 model in the previous study “Detecting digital dermatitis with computer vision.” For future application, the comparison of computer vision models for a given task can help identify which approaches perform best for various domains. For instance, the R-CNN models including Faster and Cascade R-CNN can provide better detection boundaries in facilities where hooves can overlap, such as a heifer raising facility or feedlot. However, for applications where hooves are clearly separated such as a milking parlor, the YOLO models are an excellent approach for the detection of DD lesions.
Future Direction
The computer visions models for DD detection can be expanded to include more images from more farms for different on-farm implementation. The YOLO framework can classify multiple diseases simultaneously with additional hoof images. The model can be expanded to include more DD M-stage or different hoof diseases in different settings. The model requires additional images of different cattle breeds, animal species, and geographical regions for robust predictions.
Accessibility for computer vision tool can promote the growth of such techniques in veterinary medicine. The implementation can be extended to other platforms including iOS or Android application for mobile use as well as using a docker container for cloud-based devices. The application of such methods in hand-held devices and collaboration with cattle professionals can generate a rich, diverse library of images for optimization and validation. Ultimately, the proposed computer vision tool can potentially be used to improve animal welfare and production for large-scale facilities at the herd-level.