Comparative analysis of real-time object detection algorithms for digital dermatitis in dairy cattle

Objectives

Digital dermatitis (DD) is responsible for ulcerative lesions on the interdigital space on the hoof and at the skin horn border on the coronary band of the claw. DD is associated with massive herd outbreaks of lameness and influences welfare and production. Early detection can lead to prompt treatment, reduce costs, and decrease lameness. The study aims to build an application for the real-time detection of DD in dairy cows. We compared various computer vision (CV) models for speed using inference time and for scoring using various performance metrics. The best model was automated for detection using downloaded and livestreaming video. We hope the tool can help to minimize the adverse effects of DD-associated lameness in all cattle by means of early detection, prevention, and prompt treatments.

Materials and methods

Images were collected from commercial dairy farms facing the rear foot of the interdigital space of the hoof. Images were scored by a trained investigator using the M-stage DD classification system. The classification system describes various clinical stages over the course of the disease based on morphological observations between healthy, active, healing, and chronic stages. A primary library of a single hoof per image and two class labels includes 1,177 M0/M4 and 1,050 M2 images. A secondary library of multiple hooves per image and five labels consists of 240 M0, 17 M2, 51 M2P, 114 M4H, and 108 M4P labels for a total of 409 images. Faster region- based convolutional neural network (R-CNN), Cascade R-CNN, Single Shot MultiBox Detector (SSD), and SSD Lite in addition to the four latest versions of You Only Look Once (YOLO): YOLOv3, Tiny YOLOv3, YOLOv4, and Tiny YOLOv4 were all trained to detect M-stages of DD for the respective datasets. Frames per second (FPS) were used for inference time to evaluate the time to predict the bounding box and class labels of the located objects in an image. In addition, precision, recall, and mean average precision (mAP) at intersection over union (IOU) of 0.5 were used for performance measures to compare between the predictions made by the CV models and a trained investigator (ground truth). [...]