Attention Guided Deep Neural Network for Animal Ear Tag Classification in Low-Resolution Images

Thinh Pham-Duc, Mohib Ullah, Muhammad Mudassar Yamin, Adane Nega Tarekegn, Azeddine Beghdadi, Faouzi Alaya Cheikh, Habib Ullah

Abstract: We developed a deep learning method with attention features for identifying ear tags of animals in the livestock sector. The model consists of channel and spatial attention that performs channel-wise feature calibration for improving the model representation power. We investigated the impact of ear-tag crop size from the original frame to achieve desirable efficiency and reduce the image’s noise and background. To validate our approach, our model is trained on the data collected at an animal farm in Norway, which the experts manually annotated. The model results are presented according to two performance metrics, accuracy and the F-score. An improvement of nearly 10% is obtained by our model over the state-of-the-art method. We also measured the time complexity and conducted an ablation study to demonstrate the effectiveness of our model.