Analysis of Human Ear Recognition Algorithms Based on Pseudo-Labelled Semi-Supervised Strategies

Abstract

This article focuses on human ear identification technology. A recognition algorithm based on Pseudo Labeling semi supervised learning is proposed to address the problem of insufficient ear image annotation data. This algorithm can effectively utilize unlabeled data to reinforce model training, significantly improving the detection performance of small targets and tail classes. Based on the experimental results of the baseline model, this paper ultimately chooses the Faster R-CNN model and combines it with the Feature Pyramid Network (FPN) to generate feature vectors using Global Average Pooling (GAP), thus achieving an end-to-end human ear identity recognition process. The innovation of this article lies in the improvement of the MeanTeacher algorithm process, while introducing two data augmentation techniques: pixel level mixed pseudo label Mixup and image stitching pseudo label Mosaic. Among them, Mixup reduces the negative impact of missed targets, while Mosaic further enhances the model’s recognition ability for small targets by increasing the number of labels for small-scale targets. These two techniques work together with the improved algorithm flow to enhance the performance of deep convolutional neural networks in semi supervised object detection tasks. All experiments were conducted on high-performance servers. By comparing the baselines of different models, this article selects Faster R-CNN as the best model and makes targeted improve- ments and optimizations to it. Although the generalization ability and robustness of the model still need further improvement, this study has opened up a new path for the ear recognition technology, which has guiding significance for future algorithm improvement and application expansion.