Separable Confident Transductive Learning for Dairy Cows Teat-End Condition Classification

SIMPLE SUMMARY: The health of dairy cows is important for milk quality and the health of the mammary gland. Traditionally, teat-end health has been assessed manually through visual inspection of teat-end callosity thickness and roughness (i.e., hyperkeratosis), which is a risk-factor for mastitis. Here, we describe a computer-vision approach to replace the time-consuming and expensive manual assessment of teat-end hyperkeratosis. Using separable confident transductive learning, a convolutional neural network is trained with the goal of increasing the feature differences in the images of teat-ends with different classifications of hyperkeratosis. When compared with the traditional approach of transfer learning of a convolution neural network for classifying the extent of hyperkeratosis, the overall accuracy of our model increased from 61.8 to 77.6%. This substantial improvement in accuracy renders the possibility of using image-based machine learning to routinely monitor hyperkeratosis on commercial dairy farm settings. ABSTRACT: Teat-end health assessments are crucial to maintain milk quality and dairy cow health. One approach to automate teat-end health assessments is by using a convolutional neural network to classify the magnitude of teat-end alterations based on digital images. This approach has been demonstrated as feasible with GoogLeNet but there remains a number of challenges, such as low performance and comparing performance with different ImageNet models. In this paper, we present a separable confident transductive learning (SCTL) model to improve the performance of teat-end image classification. First, we propose a separation loss to ameliorate the inter-class dispersion. Second, we generate high confident pseudo labels to optimize the network. We further employ transductive learning to narrow the gap between training and test datasets with categorical maximum mean discrepancy loss. Experimental results demonstrate that the proposed SCTL model consistently achieves higher accuracy across all seventeen different ImageNet models when compared with retraining of original approaches.