Research Of Dehazing Algorithm Based On Image Restoration And Deep Learning

With the rapid development of science and technology,the demand for intelligence in human society is becoming more and more extensive and involves various fields.Computer vision is an important branch in the field of artificial intelligence,which is the simulation of biological vision by computer devices to achieve understanding and processing of image tasks.Currently,computer vision plays an important role in application situations such as surveillance systems,payment security and autonomous driving.However,some computer vision systems in outdoor conditions are susceptible to adverse weather conditions that can cause such systems to degrade or even fail.The most common severe weather is caused by large amounts of water vapor,dust and other suspended particles in the air.Under these conditions,the quality of the acquired images is severely degraded,color fidelity is lost,and clarity is reduced because of the scattering and refraction of light reaching the imaging device.Hazy images have problems such as blurred targets,loss of detail,and color distortion,which have a detrimental effect on subsequent computer vision tasks.Therefore,the image clarification task in haze scenes has important research value and significance.This thesis analyzes the current mainstream dehazing algorithms in detail.Combining the technical routes of existing algorithms,improvement strategies are proposed from two directions,image recovery and deep learning,respectively.Based on image restoration,a novel defogging algorithm is proposed by analyzing the physical degradation model of hazy scene images and the inherent defects of such methods.(1)Aiming at solving the problem of contour effect and color distortion in the sky area of the dark channel prior algorithm,a dehazing algorithm for sky area segmentation and transmission mapping of different areas is proposed.Firstly,the sky area of image is roughly segmented by adaptive threshold method,and atmospheric light value is estimated in the sky area.Secondly,the dark channel is improved by combining the super pixel segmentation method to obtain initial transmission,and refined transmission is obtained by the guided filtering method.Adaptive threshold segmentation is performed on the refined transmission,and the largest connected domain is retained to achieve fine segmentation of the sky area.Finally,different transmission mapping methods are proposed for the sky and non-sky areas to obtain the final transmission.The atmospheric scattering model is used to restore the image.The experimental results show that the restored image performs well in both subjective vision and objective indicators,which effectively solves the defect that the dark channel prior algorithm is easy to fail in the sky area,and can restore a more natural sky and weaken the halo effect in the edge area.Based on deep learning,two dehazing network models are proposed in the directions of indirect-type intermediate feature estimation structure and end-to-end structure,respectively.(2)To address the problems such as image quality impairment and contrast degradation in hazy conditions,a dehazing model based on attention and multi-scale grouped parallel convolutional neural networks is proposed on the basis of intermediate feature structure dehazing.Firstly,to avoid the error amplification problem caused by separate learning,the atmospheric scattering model is transformed to facilitate the focused learning of transmission and atmospheric light.Among them,the multi-scale convolution structure can be used to extract richer multi-level features of haze images.Also,in order to prevent problems such as image information loss and gradient disappearance,the residual structure is added to the parallel layers separately,so that the training results contain more content information and texture structure.Finally,a clear image is obtained by combining the improved restoration model.Experiments show that the proposed network model has satisfactory dehazing effect for both synthetic haze images and real haze images,also achieves satisfactory results in objective evaluation.(3)On the basis of end-to-end structural dehazing architecture,existing deep learningbased dehazing methods are basically analyzed using convolutional neural networks(CNNs).Due to the inherent characteristics of CNNs,it has limited ability to express correlations between image information.So,general CNN-based dehazing models tend to have complex structures and poor robustness.To solve the above problems,a new network combining CNN and Transformer architecture is proposed: Multi-Scale Transformer Fusion Dehazing Network(MSTFDN)to achieve end-to-end dehazing.MSTFDN consists of three different functional modules: multi-scale Transformer fusion module(MSTFM),feature enhancement module(FEM)and color recovery module(CRM).Among them,MSTFM consists entirely of a multiscale Transformer block structure for capturing the long-range dependence of image information in space;FEM is used to enhance the forward features and extract features at different depths;CRM is used to obtain the final clear image and recover its color.The effectiveness of each module is demonstrated by corresponding ablation studies and the best multi-scale Transformer combination is selected.Extensive experiments on synthetic and real haze image datasets show that the proposed model is robust,outperforms state-of-the-art methods in qualitative comparisons,and performs well in quantitative comparisons.