Lightweight Image Dehazing Neural Network Model Based on Estimating Medium Transmission Map by Intensity

Abstract

AbstractSingle image dehazing is a challenging ill-posed problem. The key to achieve haze removal is to estimate an accurate medium transmission map. By redefining the atmospheric scattering model, we obtain a new transmittance map scattering model, haze image and haze-free image, derive the medium transmission as a function of the scene intensity only, also deduce a priori condition that the intensity of hazy image is higher than that of haze-free image, and propose a lightweight image dehazing neural network (Intensity neural network, I-Net) based on estimating medium transmission map by intensity. I-Net uses a convolutional neural network (CNN) as the backbone, and takes the intensity of hazy image as the input, and outputs the intensity of haze-free image, the medium transmission map and the original haze-free image, also joints the priori condition derived previously to obtain a more accurate medium transmittance map. In this paper, the dehazing algorithm obtains the intensity of haze-free image through I-Net, derives the transmittance map using the functional relationship between transmittance and scene intensity, and finally recovers the original haze-free image through the transmittance map scattering model. The experimental results show that our dehazing results are clearer and more natural. The subjective and objective evaluations show that our image dehazing algorithm can achieve better dehazing results than traditional algorithms, and outperforms current advanced algorithms in terms of Peak Signal to Noise Rate (PSNR) and Structure Similarity Index Measurement (SSIM).KeywordsSingle image dehazingAtmospheric scattering modelMedium transmission mapDeep CNN