改进生成对抗网络的雾霾天气交通标志识别算法

The algorithm for traffic sign recognition in hazy weather is enhanced by incorporating an improved generative adversarial network

为提高雾霾天气下交通标志识别的准确性,提出一种改进生成对抗网络(GAN)的雾霾天气交通标志识别算法。该算法主要分为2个部分:第1部分是多尺度GAN的图像去雾,在生成器中增加多尺度卷积和感知损失函数,多尺度卷积有利于提取特征,感知损失可以在图像的深度特征上保留内容、风格等高级语义信息,使去雾效果更符合人眼对图像质量的感受;第2部分是交通标志识别,在原有YOLOX-S模型的基础上增加感受野更小的160×160检测层来降低小尺度交通标志的漏检率,其次在主干网络中增加坐标注意力(CA)机制强化特征网络。实验结果表明,提出的去雾模型具有较好的效果,评价指标PSNR和SSIM的结果都优于其他代表性算法;交通标志识别算法与原模型进行比较,准确率、mAP值、召回率分别提高了2%、4%和7%,验证了模型的有效性。

The recognition of traffic signs,crucial for ensuring standardized vehicle driving,is significantly impaired by smogs.To enhance the accuracy of traffic sign recognition in smoggy weather,this paper proposes an algorithm based on an improved Generative Adversarial Network(GAN)specifically designed for smoggy scenarios.The algorithm consists of two main parts:first,a multi-scale GAN is employed to defog images by incorporating multi-scale convolution and perceptual loss functions into the generator.Multi-scale convolution facilitates feature extraction while perceptual loss preserves high-level semantic information like content and style in depth features,resulting in defogging effects that align better with human visual perception.Second,for traffic sign recognition,a smaller 160×160 detection layer with a reduced receptive field is added to the original YOLOX-S model to minimize missed detection of small-scale traffic signs.Meanwhile,a coordinate attention(CA)mechanism is introduced into the backbone network to strengthen feature representation.Our experimental results demonstrate the proposed defogging model outperforms other representative algorithms in such evaluation indices as PSNR and SSIM.Moreover,compared to the original model,our traffic sign recognition algorithm improves accuracy by 2%,mAP value by 4%,and recall rate by 7%,demonstrating its fair effectiveness.