Research On Image Dehazing Algorithm

Images captured by optical sensors in foggy or haze scenes will be scattered by atmospheric particles,resulting in loss of image details,dim colors,and reduced brightness.The contrast and color fidelity of images will be reduced,which directly affects people's visual perception for images and the performance of the computer vision system,such as seriously reducing the accuracy of the target recognition and classification by the monitor.Therefore,image defogging is a very significant task.From the early dark channel prior defogging algorithm to the recent image defogging algorithm based on deep learning,the defogging effect is not ideal in images with sky.Fog maps are almost always obtained outdoors,and outdoor images often contain the sky,so image defogging algorithms for the sky area must be researched.This paper proposes two defogging algorithms based on the principle of dark channels prior that can process images with sky areas.The related works are as follows:1.A new sky segmentation algorithm is proposed,which can accurately segment the sky and non-sky areas(including the sky in a grid-like sparse area).Secondly,the atmospheric light value is obtained in the obtained accurate sky area.We calculate different transmittance in the sky area,refine them with guided filtering,and improve the visual effect of the restored image with improved adaptive tone mapping.2.A sky recognition algorithm based on Fisher criterion function is proposed to separate the sky and non-sky regions.The sky and non-sky are used with different transmittance,and then the atmospheric scattering model is used to obtain the restored image.In order to make the defogged image have more natural colors,we also designed an adjustment formula for the saturation component,and constructed an improved gamma correction algorithm to improve the image contrast.Image defogging and image Super-Resolution are both important image tasks in computer vision,so this paper also has additional research on image Super-Resolutionalgorithms.The existing deep convolution neural networks almost have huge parameters and the overall calculation volume,so designing a lightweight image super-resolution network to adapt the limited hardware level has been becoming the trend.This paper proposes a kind of relatively lightweight deep network.The related works are as follows:Two lightweight criteria for image super-resolution networks are proposed;an image super-resolution network with deep aggregation structure with fewer parameters is designed.