Self-Supervised Learning for Swine Ear-Tag Classification with Attention Module Model

We proposed an attention-based deep framework for animal ear-tag classification in the livestock industry. In addition to spatial and channel attention, the model incorporated squeeze excitation block that performed channel-wise feature calibration to improve the model representation power. Moreover, this backbone was transferred to self-learning model on unlabeled data - Barlow Twins and combined with focal loss. We investigated the impact of ear-tag crop size from the original frame to achieve desirable efficiency and to reduce noise in the images, such as the pen or swine. The model was trained on manually annotated data collected at a swine farm in Norway to validate our approach. The results of model were presented in terms of two performance metrics: accuracy and F-score. A comprehensive comparison with state-of-the-art model was conducted, and more than 10% improvement was achieved. The ablation study was also performed and proved the effectiveness of our approach.