基于深度学习的序列交通图像去雾方法

Dehazing Method for Sequential Traffic Image Based on Deep Learning

受大气散射影响,序列交通图像的细节特征损失较大,图像像素较低,为此提出基于深度学习的序列交通图像去雾方法。根据雾化图像表达式和相同大气密度下光线载体传输图定义式,构建大气散射模型,运用该模型获取图像特征序列。利用深度学习自编码网络中的网络层,建立输出定义式与特征损失函数式,在样本中引入图像特征序列,得到特征块序列,依据雾特征图和散射率的非线性映射关系获取散射率图。采用修正函数调整卷积层输出像素值为正,将局部块代入自编码网络,并添加特征块序列至卷积神经网络的输入层,对输出的散射率图进行导向滤波处理,实现序列图像雾特征的去除。仿真结果表明,去雾后图像的细节特征更加突出,大幅度提升了图像的对比度。

The atmospheric scattering leads to the loss of detail features of the sequential traffic image, and the image pixels are too low. Therefore, a defogging method for sequential traffic image based on deep learning was proposed. According to the expression of foggy image and the definition of light carrier transmission diagram under the same atmospheric density, the atmospheric scattering model was constructed, and then the model was used to find out the sequence of image features. Moreover, the network layers in self-coding network based on deep learning were used to establish the output definition and the feature loss function. The image feature sequence was introduced into the sample to obtain the sequence of feature blocks. According to the nonlinear mapping relationship between the fog feature map and the scattering rate, the map of scattering rate was obtained. In addition, the output pixel value of convolution layer was adjusted to be positive by the correction function. Meanwhile, the sequence of feature blocks was added to the input layer in convolutional neural network, and the output scatter rate map was guided and filtered. Finally, the fog feature of sequence image was removed. Simulation results show that the detail features of image are more prominent after defogging, and this method greatly improves the contrast.