Cow Re-identification with Attention Branch Network and Its Evaluation for Real-World Applications

Abstract

In recent years, the number of dairy farms has been decreasing in Japan. In order to reduce the burden on dairy farmers required for early detection of disease signs, methods have been developed for identifying individual dairy cows using image recognition based on machine learning. One of the problems in the conventional methods is that they require much time and effort to prepare training data every time dairy cows are replaced due to the entry of new dairy cows, etc., and it takes time and effort to create training data. In addition, since the judgment process is black-boxed, another problem is that it is difficult to interpret the result. In this paper, we introduce Attention Branch Network to generate visual explanations for individual identification and to improve identification accuracy. Furthermore, to clarify the effect of dairy cow replacement on the identification accuracy, assuming a situation close to practical use, we evaluated the accuracy while changing the proportion of labels not included in the training data.