Research On Enhancement And Speckle Reconstruction Of Scattering Degraded Image

Image degradation caused by scattering is unavoidable in optical imaging.According to the intensity of scattering when photons are transmitted in different scattering environments,the image quality can be divided into “visible but not clear” weak scattering image degradation and “completely invisible” strong scattering image degradation.For the degradation problem in the weak scattering environment such as haze,image processing based haze removal method is the current development trend to improve the image quality;for the degradation problem in the strong scattering environment such as hairy glass,speckle image reconstruction based on deep learning method is an important part of computational imaging research.This paper studies how to recover high-quality target objects from scattered light field and carry out image dehazing and speckle reconstruction,mainly including the following aspects:In order to solve the problem of haze caused by atmospheric disturbance in visible light imaging,a filtering method based on global frequency domain-aware was proposed.The method adjusts the frequency domain filter shape according to the low frequency component representing image energy,the medium frequency component representing image detail and the high frequency component representing image noise,and enhances image detail while suppressing the noise of the original haze image.The PSNR value of the dehazed image reaches 1.15 times and the SSIM value reaches 1.17 times of the state-of-the-art methods.Meanwhile,since the method performs global filtering in the frequency domain,the computational efficiency can be significantly improved.Aiming at the difficulty of visible imaging to solve the haze problem in distant scenes,sky and other regions,combined with near-infrared imaging technology,a method for dehazing by fusion of near-infrared and visible is proposed and an energy conservation model of the image is constructed to allocate the proportion of the energy of the nearinfrared image and the visible image to the fusion result,so that the color of the fusion result is evenly distributed.On this basis,a color mapping method is proposed to ensure that the final result has rich color information.The method can make the fusion result with high color contrast and detail visibility,and the overall image fusion quality is effectively improved compared with the state-of-the-art fusion methods tested under the public haze scene dataset.In order to solve the problem of low resolution in speckle image reconstruction by deep learning methods,two network cascade models are proposed.On the one hand,for simple data sets with small samples,a two-stage self-back fine network is proposed to improve the image reconstruction with higher resolution by repeatedly mining the effective utilization of the original data,which improves the average reconstruction accuracy by 18.3% compared with the single-stage network;on the other hand,for the problems of missing and misplaced information in the reconstruction results,an objective and perceptual dual loss function channel fine network framework is proposed to first reconstruct high and low-dimensional feature images separately,and achieve more accurate reconstruction of edge information by fusing the complementary information of the two channels,and the average reconstruction accuracy is improved by 3.2% compared with that before fusion.To address the problem of detail loss and the distortion of the target in the reconstruction results of the speckle image,an optimization method of maximizing the effective receptive field of the network is proposed;and two different loss functions,objective and perceptual,are still used to obtain the complementarity information during speckle reconstruction,the advantages of the two reconstruction results of the first stage are automatically learned through a two-stage augmented network.After a series of experiments,the overall average error rate of the reconstruction results is reduced by 14.1%on the Fashion-MNIST dataset compared with the state-of-the-art algorithms.In general,the image dehazing and speckle reconstruction methods studied in this paper can improve the visibility of haze images in weak scattering environments,and can achieve more refined reconstruction of speckle images in strong scattering environments.The proposed algorithms and network models can be used as a reference for research on computational imaging work in scattering degradation environments.