基于自注意力模型的图像去雾算法

Image defogging algorithm based on self-attention model

为了提高路侧相机在雾天气象环境下的图像对比度和清晰度,基于自注意力模型建立了一种图像去雾算法。通过视觉感知实时获取道路交通目标信息,是实现道路数字化与智慧化的基础,而雾天环境会使感知算法出现性能衰退。文中在生成式对抗神经网络框架下建立了基于自注意力模型的图像去雾算法,能够有效建模图像中不同区域间的关系,同时解决实际环境中难以建立无雾-有雾图像对训练集问题。利用Vision-Transformer网络提取图像中雾层特征,通过U-Net网络对图像雾层进行估计;采用卷积神经网络对去雾图像进行评估,通过对抗训练方式对网络参数进行优化;基于实际高速公路路侧相机采集了自然雾天环境下的道路交通图像数据,并结合Foggy Driving和O-HAZE开源数据集对提出的图像去雾算法进行了验证,结合主观和量化指标对算法的去雾效果进行了验证评估,表明该算法能够有效提高图像质量。

In order to improve the sharpness and contrast of the images obtained by roadside cameras in foggy days,an image dehazing algorithm is proposed on the basis of the self-attention model. To collect the road traffic target information in real time by visual perception is the basis of road digitalization and intelligence. However,the environment in foggy days will make the performance of the perception algorithm declined. In this paper,an image dehazing algorithm based on self-attention model is established under the framework of generative adversarial neural network,which can model the relationship among different areas of the image effectively and solve the problem that it is difficult to establish the training set of fog-free-foggy image pairs in the actual environment. The Vision-Transformer network is used to extract the features of fog layer in the image. The fog layer in the image is estimated by means of the U-Net network. The convolutional neural network(CNN)is used to evaluate the dehazing image. The means of adversarial training is used to optimize the network parameters. The image data of road traffic in foggy days is collected on the basis of the cameras on roadside of the highway. The proposed image dehazing algorithm is verified in combination with the open source data sets Foggy Driving and O-HAZE,and the defogging effect of the algorithm is verified and evaluated in combination with subjective and quantitative indicators. The results show that the algorithm can improve the image quality effectively.