The Research Of Image Dehazing Algorithm Based On Deep Learning

Due to the scattering of atmospheric particles in haze weather,the images obtained through video and image acquisition systems usually have low visibility,which will seriously affect subsequent performance of target recognition and target tracking by computer vision systems.At present,image dehazing algorithms can be divided into traditional dehazing algorithms and dehazing algorithms based on deep learning.Traditional dehazing algorithms generally use image enhancement to enhance the contrast of foggy images,or use prior theory to estimate the transmittance values and atmospheric illumination values of foggy images,and then restore fog-free images through atmospheric scattering models.Dehazing algorithms based on deep learning are generally modeled based on the prior information of the foggy image,and then the transmission or non-fog map of the foggy image is estimated by supervised learning or unsupervised learning.These algorithms are generally more general and more efficient.This paper mainly studies the image dehazing algorithm based on deep learning.The main contributions are as follows:(1)In-depth analysis of some typical dehazing algorithms in the field of traditional image dehazing,and performance comparison with image dehazing algorithms based on deep learning.This paper conducts experiments on the classic image dehazing datasets RESIDE and O-HAZE,and uses the no-reference image quality evaluation indicators and running time to measure the dehazing effect of many typical image dehazing algorithms.Experimental results show that the dehazing algorithm based on deep learning not only has good dehazing effect,but also has higher dehazing efficiency on single image.(2)Aiming at the problem that the current dehazing network model has too many training parameters and affects the dehazing efficiency of the image,this paper proposes a new fast dehazing model called FAOD-Net based on lightweight network for dehazing single image.FAOD-Net model is based on a lightweight architecture that uses depthwise separable convolutions to build lightweight convolutional neural networks.In addition,this paper adds a pyramid scene parsing network to FAOD-Net model to aggregate the context information of different regions of the image,thereby improving the ability of the network model to obtain the global information of the foggy image.This paper uses the RESIDE training set to train the FAOD-Net model,and conducts extensive experiments on the RESIDE test set,using full-reference and no-reference image quality evaluation indicators to measure the effect of dehazing.The experimental results show that the FAOD-Net model has satisfactory results in terms of dehazing effect and speed.(3)Aiming at the problem that the current image dehazing algorithm is likely to cause color distortion in the defogged images,this paper proposes a new image dehazing model based on the color feature extraction convolutional network called CIASM-Net.CIASM-Net model includes color feature extraction convolutional network and deep dehazing convolutional network.Among them,the color feature extraction convolutional network is used to extract the features of the RGB color space of the foggy image.The deep dehazing convolutional network improves the inverse atmospheric scattering model convolutional network IASM-Net and uses a multi-scale convolution layer to estimate the transmittance map.In addition,this paper adds a pyramid scene parsing network to the CIASM-Net model to extract global features.This paper uses the classic RESIDE training set to train the network model.Experimental results on the RESIDE test set prove that the CIASM-Net model has satisfactory dehazing effect.