Self-supervised Siamese Networks with Squeeze-Excitation Attention for Ear Image Recognition

As a new biometric technology, ear recognition is widely used in various fields such as individual identity verification, security monitoring, and access control due to its advantages of uniqueness, stability, and ease of acquisition. To deal with the challenges of labeled data scarcity and the inherent variability in ear shapes, which hinder the performance of traditional ear image recognition systems, a novel ear recognition model based on Siamese Network with Squeeze-Excitation Attention (SE-SiamNet) is proposed, which takes the pretrained ResNet50 network as backbone, and takes the Softmax classifier with cross-entropy loss function as classification head. SE-SiamNet utilizes the principles of self-supervised learning, where the net-work is trained to predict the similarity between pairs of ear images, where in each training batch, the nearest image is used as positive sample, and the furthest image is used as negative sample. Therefore, the meaningful features are extracted from the data without explicit supervision. To enhance the feature discriminative capability of the model, squeeze-excitation attention mechanism was integrated into the backbone, which allows the network to selectively emphasize important features within the ear images. By focusing on the most discriminative parts of the ear images, the model becomes more robust to variations in shape, size, and orientation. Results from three publicly available test datasets on the Kaggle platform demonstrate that, compared to classical convolutional networks, the SE-SiamNet model achieves an accuracy, precision, and recall rate of over 95%, validating that SE-SiamNet is more reliable and data-efficient for developing biometric systems without relying on labeled training data.