A Haze Removal Method via The Fusion of Gaussian Low-Frequency Multi-Scale and Median Rank Detail Perspective Network

Images captured outside in bad weather are polluted by colloidal particles and atmospheric droplets.These acquired pictures are the source of major mistakes in digital image vision systems because they are prone to low contrast, poor visibility, and color distortion.As a result, defogging research is important for real-world applications.This study uses a fog image degradation model to define picture defogging as a mathematical inversion and image restoration procedure.The global atmospheric light A and transmittance can be precisely estimated by combining a Gaussian low-frequency multi-scale convolutional network with a median rank detail perspective network (GLFM-MRDP Net).A Gaussian low-frequency multi-scale convolutional network is first used to obtain the low-frequency part of the image, and then the block convolution model is adopted to acquire accurate A. Then the feature extraction subnet and median rank detail optimized network is adopted to obtain transmittance, which can suppress image noise while retaining details as much as possible and extracting features through convolution.Comparative experimental findings demonstrate that our approach is successful in handling dense fog, complicated scenes, and multi-scale pictures.Besides, the no-reference and full-reference evaluation metrics of our method are superior to methods [8,11,17,22].As a result, our technology outperforms four other cutting-edge defogging techniques in terms of aesthetic effect, applicability, and running speed.