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

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Traffic sign recognition is vital for autonomous vehicles and traffic safety.
• Challenges include high intraclass variability, interclass similarity, and complex backgrounds.

Purpose of the Study:

• To propose a novel invariant cue-aware feature concentration transformer (TTSNet) for effective traffic sign recognition.
• To address the limitations of existing methods in handling visual feature variations and background complexity.

Main Methods:

• Introduced three novel modules: attention-based internal scale feature interaction (DLFL), cross-scale cross-space feature modulation (SSFM), and eliminating spatial and information redundancy (ESIR).
• DLFL extracts invariant cues using discriminative value-based feature selection.
• SSFM aggregates multi-scale features, while ESIR reduces spatial and channel redundancy for improved representation.

Main Results:

• TTSNet achieved state-of-the-art performance on benchmark datasets.
• Achieved 89.1% accuracy on the T100K dataset.
• Achieved 89.97% accuracy on the CTSDB dataset.

Conclusions:

• The proposed TTSNet effectively learns invariant and core information from traffic signs.
• The novel modules significantly enhance feature representation and recognition accuracy.
• TTSNet demonstrates superior performance in complex traffic sign recognition scenarios.