Dynamic threshold semi-supervised traffic sign classification based on residual convolutional autoencoder and deformable convolution

Abstract

Traffic sign classification is a vital part of intelligent driving, and convolutional classification methods have been widely applied. However, traffic sign images captured in motion are prone to distortion and noise interference. Coupled with the scarcity of labeled images, these factors collectively impair model classification accuracy. To address the above problems, this paper proposes a dynamic threshold semi-supervised image classification method based on FixMatch, which combines residual convolutional autoencoder (RCAE) and deformable convolution. Firstly, labeled images are used to train the supervised denoising RCAE for feature extraction and label prediction of noisy images. Afterward, the attention mechanism selects significant features and locates key regions in original images. Deformable convolution is applied on these regions to effectively guide offset learning. Fusing high-importance features with those extracted by deformable convolution serves as auxiliary information to compensate for insufficient global information extraction, ultimately enhancing image classification performance. Additionally, to overcome labeled data scarcity, we propose a dynamic threshold screening strategy and a weighted label fusion method based on FixMatch, which improve the accuracy and utilization of pseudo-labels. Experimental results on the GTSRB and BelgiumTS datasets demonstrate a significant improvement in classification accuracy.