Chinese Traffic Guide Panel Text Detection Based on Pixel Aggregation

The detection and recognition of road traffic signs has always been a key research project in the fields of assisted driving and intelligent transportation, which has ignited the enthusiasm of many researchers. However, most methods focus on the study of symbolic traffic signs, the study of detecting and recognizing textual traffic signs is still challenging. In order to realize the efficient and accurate detection of text in Chinese traffic guide panels, a text detection network based on pixel aggregation (PA) is proposed, which can detect text kernels and precisely aggregate their surrounding text pixels using similarity vectors. More specifically, the model neck is composed of one Single-scale Transformer Encoder (STE) and one Cross-scale Feature Fusion Module (CFFM). STE can explore the position-aware feature representation with self-attention mechanism. CFFM is a light-weight U-shaped module, which can merge multi-level information to guide better segmentation. After CFFM, the features from different levels are gathered into a final feature for segmentation. Experimental results on CTST-1600 show that our proposed method outperforms other text detection methods while ensuring real-time performance.