TTSNet: Traffic sign recognition via a transformer by Learning Spectrogram Structural Features

Traffic sign recognition is crucial not only for autonomous vehicles and traffic safety research but also for multimedia processing and computer vision tasks. However, traffic sign recognition faces several challenges, such as high intraclass variability and interclass similarity in visual features and background complexity. We propose a novel invariant cue-aware feature concentration transformer (TTSNet) to effectively address these challenges. TTSNet aims to learn the invariant and core information of traffic signs. To this end, we introduce three new modules to learn the features of traffic signs: attention-based internal scale feature interaction (DLFL), cross-scale cross-space feature modulation (SSFM), and eliminating spatial and information redundancy (ESIR) modules. The DLFL module extracts invariant cues from traffic signs through feature selection based on discriminative values. The SSFM-Fusion module aggregates multi-scale information from traffic sign images by concatenating multi-layer features. The ESIR module improves feature representation by eliminating spatial and channel information redundancy. Extensive experiments showed that TTSNet achieves state-of-the-art performance on the T100K (89.1%) and CTSDB (89.97%) datasets.