Deep learning-based model to classify mastitis in Holstein dairy cows

Abstract

The occurrence and prevalence of dairy cow mastitis has brought significant challenges to animal welfare and economy. To overcome the complexities and accumulated errors present in previous detection methods, a rapid and accurate mastitis detection approach is developed based on image processing and deep learning, leveraging thermal infrared imaging. Image processing techniques, including the Hough transform and morphological operations, are used to classify affected cows from thermal images. An image pyramid is constructed based on upsampling to tackle motion blur induced by the cows' rapid movement. The multi-scale convolution and the spatial and channel Squeeze & Excitation (scSE) block were integrated into the DenseNet-201 architecture to enhance the feature extraction process. This enabled the network to adaptively recalibrate channel-wise feature responses and strengthening the discriminative power of the learned representations. For mastitis detection, a deep learning model, the multi-scale scSE-DenseNet-201 (MS-scSE-DenseNet-201) architecture, is refined to predict the severity of mastitis. The framework takes images of both sides of the cow's udder as input, and outputs one of three mastitis severity levels: negative (N), subclinical mastitis (SCM), or clinical mastitis (CM). To assess the model's performance in detecting mastitis, a dataset comprising 5000 thermal images from 802 cows, was used. The model achieved accuracy, precision, and recall of 90.18%, 92.16%, and 88.38%, respectively, showing notable improvement over previous methods. This work integrated object segmentation and blind deblurring to strengthen the MS-scSE-DenseNet-201 in the automatic detection of cow mastitis, which will open a promising application horizon for other animal disease diagnostics.