An Improved Image Defogging Algorithm Based on Dehazenet Network

The image dehaze technology is particularly important. This paper adopts the traditional physical model and deep learning network model approach, using the improved dehazenet network, using three different scales of convolution kernel for feature extraction of the original foggy image, and in the feature learning process using 3×3 tiny convolutional kernel to increase the model’s depth; in order to avoid the network overfitting as well as to enhance the convergence speed of the network, the batch after convolution was added Batch Normalization processing is added after the convolution to estimate the transmittance t , and the original bootstrap filter is improved to a weighted bootstrap filter so as to better strengthen the edge information, and the quadtree method is used to estimate the atmospheric light value A instead of the original method to ensure a more accurate estimation of the atmospheric light. Using RESIDE dataset, the outdoor image with sky, the building image and the indoor restaurant image are used for the defogging experiments and the three values of PSRN, MSE and SSIM are evaluated.