Traffic Sign Detection and Recognition Using Multi-Scale Fusion and Prime Sample Attention

Traffic sign detection, though one of the key technologies in intelligent transportation, still has bottleneck accuracy due to small size and diversity traffic signs. To solve this problem, we proposed a two-stage CNN object detection algorithm based on multi-scale feature fusion prime sample attention. We improved original Faster R-cnn model terms extraction sampling strategy. For extraction, elevate ability neural networks detect objects, adopted HRNet as extractor. There are four stages - series high resolution subnets starting point with repeated adding parallel low form other stages. In whole process, information multi-resolution sub-network is repeatedly exchanged perform fusion. strategy, simple effective learning strategy called Prime Sample Attention (PISA), consisting Importance-based Reweighting (ISR) Classification Aware Regression Loss (CARL). PISA concepts IoU Hierarchical Partial Sorting (IoU-HLR) Score (Score-HLR), which sort importance positive samples negative mini-batch respectively. With method, training process focusing rather than evenly treat all ones. The complexity our method lower that state-of-the-art. After experiments by TT100K dataset, can attain comparable or even better robustness.