Automatic traffic-sign detection is a hot topic in computer vision and one of the critical technologies of intelligent transportation. The Transformer structure has recently become a research hotspot due to its excellent performance. We hope to apply this structure to the design of traffic sign detection algorithms. Therefore, we make some improvements to Sparse R-cnn, a neural network model inspired by Transformer. Sparse R-cnn is a novel model, and its core idea is to replace hundreds of thousands of candidate anchors in the RPN network with a small set of proposal boxes. The experiments in our paper proved that the performance of the Sparse R-cnn model is better than other existing general object detection models. Based on the original Sparse R-cnn inspiration, an improved Sparse R-cnn model is proposed. First, a novel backbone for the task of traffic-sign detection is proposed. Multi-scale fusion structure is the essential method of improving the algorithm for small target detection, so improving the multi-scale capability of the backbone is a required method for designing traffic sign detection. So, we made further improvements to the existing backbone ResNest. We enhanced the multi-scale representation ability of the backbone by constructing hierarchical residual-like connections within each single radix block in the original ResNest. We call the improved backbone Res2Nest. The novel backbone proposed by us shows better performance without introducing excessive computational costs to the model. In addition, the attention mechanism is also an effective method to improve the detection of traffic signs, so we set up a branch network for recalibrating the channel feature response adaptively through the Global Average Pooling (GAP) operation and a fully connected layer. It can also be seen as the implementation of the crosschannel self-attention mechanism. After experiments by TT100K dataset, our method would attain a better accuracy and robustness.

No abstract