High-similarity sheep face recognition method based on a Siamese network with fewer training samples

Abstract

The accurate recognition of sheep identity is significant for sheep management, breeding, production, and the traceability of meat products. Currently, deep learning technology provides a method to recognize the corresponding identity of sheep by their facial features. However, some sheep faces of the same breed are highly similar, which makes the sheep face recognition models based on shallow feature extraction networks face challenges. To address these challenges, this study proposes a high-similarity sheep face recognition model based on a Siamese network, named Siamese-High-Similarity Sheep Face Recognition (Siamese-HSFR). Siamese-HSFR uses contrastive learning to determine the similarity between a pair of sheep face images, assessing the probability that they belong to the same sheep. In the feature extraction network of Siamese-HSFR, two extraction modules are introduced, namely Residual Fusion Block (RF_Block) and Enhanced Identity Block (EI_Block), aiming to extract more detailed and robust sheep face features. Furthermore, by introducing the three-dimensional attention mechanism in the EI_Block, the SAM Enhancement Block (SAM_Block) is constructed to enhance the discriminative capability for high-similarity face features. Small-tailed Han sheep were taken as the recognition object in this study. Through the preliminary experiments, we found that some sheep have distinctive facial features, making them relatively easy to identify. In contrast, others lack such distinctive features, leading to high similarity in face images. To further investigate the recognition model’s performance on two types of sheep faces, two datasets were established for training, including datasets of prominent features and high-similarity sheep faces. The experimental results show that Siamese-HSFR achieved the best performance on both the sheep face datasets, with accuracy reaching 97.2 % and 90.5 %, respectively, surpassing the current state-of-the-art sheep face recognition models. On the other hand, experiments indicate that when the training set comprises only 10 % of the total data, Siamese-HSFR achieves the accuracy of 92.1 % and 86.5 % on the two groups of sheep face data, demonstrating its suitability for training under small-sample conditions. The findings of this study provide technical support for high-similarity sheep face recognition tasks and small-sample training.