Dynamic convolution-based image dehazing network

Convolutional neural networks use a convolutional kernel with static weights for processing non-uniform haze or dense fog, which may lead to redundancy of network parameters. To address the structural limitations of convolution, dynamic convolution has been proposed and received wide attention; however, its direct application to image dehazing tasks still suffers from parameter redundancy, simple structure, and lack of information exchange in different dimensions. To solve these problems, a novel dynamic convolution-based image dehazing network DyStd-Net is proposed, which uses a novel dynamic convolution TDyConv, which uses the Transformer mechanism to dynamically adjust the weights of the output channel dimension and spatial dimension of the convolution kernel according to the input, giving the convolution a larger perceptual field and better nonlinear representation. In addition, a standard deviation normalization scheme StdNorm for the dynamic convolution kernel weights is explored to accelerate the dynamic convolution training. DyStd-Net adopts a U-Net-like structure and combines dynamic convolution with depth-separable convolution, making full use of image features of different dimensions to recover fog-free images. A combination of smoothed L1 loss, SSIM loss, and Perceptual Loss is used in the training process for parameter optimization. Tests on synthetic and real fogging datasets show that DyStd-Net achieves higher PSNR and SSIM metrics and provides better subjective perception compared with mainstream dehazing algorithms.