关键词: 小目标识别；交通标志识别；膨胀卷积, 注意力机制；解耦头

Abstract:

To address the issue of decreased recognition accuracy caused by false and missed detections of small target traffic signs, a small traffic sign recognition algorithm based on CGT-YOLO is proposed. First, the Context Enhancement Module is introduced to replace the SPPF module in the YOLOv5s network, enhancing the feature representation of small traffic signs through convolutions with varying dilation rates. Next, the Global Attention Module is added after the Concat module in the YOLOv5s backbone network. The GAM extracts the CAM-enhanced features and further strengthens the global interaction between channels and spatial dimensions using 3D permutation, multi-layer perceptron, and convolutional spatial attention. This process highlights the features of small target traffic signs and mitigates the negative effects of complex backgrounds and long distances. Finally, the TSC decoupling head is utilized to separate the features for classification and localization tasks. This module generates semantically rich, low-resolution classification feature maps and high-resolution localization maps containing boundary information, effectively resolving the feature conflicts between the two tasks. Experimental results show that the improved model has improved in all indicators: the miss rate and false positive rate decreased by 12.1% and 11.6%, respectively, and the mAP(0.5:0.95) increased by 2.6%. These improvements effectively reduce false positives and false negatives, significantly, enhancing the recognition accuracy of small traffic signs.

Key words: small target recognition, traffic sign recognition, dilated convolution, attention mechanism, decoupling head