Research On Image Dehazing Algorithm Based On Deep Learning

With the development of science and technology,computer vision systems are more and more integrated into people’s daily life.Whether it is the face recognition fast cashier system that can be seen everywhere in the streets,or the new energy vehicles with a large number of cameras on the road,they all use a lot of computer vision.system.These computer vision systems are deployed in real environments,which are subject to various negative effects that differ from the ideal environment in the laboratory.Among the various negative effects,the most common ones are caused by haze weather.In such an environment,the images captured by the computer vision system will have various problems,such as the overall grayish and white phenomenon of the image.This leads to problems such as decreased contrast,poor color saturation,and reduced information contained in the image.These problems increase the difficulty of subsequent advanced computer vision processing tasks such as target detection,image classification,and semantic segmentation.Therefore,how to effectively degrade fog and blur It is particularly important to restore the image to a fog-free and clear image,that is,image defogging technology.The emergence of machine learning and deep learning has brought another possibility to the field of image defogging,and traditional image defogging algorithms are slowly being replaced.The early dehazing algorithm based on deep learning did not completely break away from the classic theory of the traditional image dehazing algorithm,but there were problems such as the dehazing effect was not obvious,and it relied too much on prior conditions.The subsequent algorithm was based on convolutional neural network and did not rely on prior conditions.Direct end-to-end output of fog-free images,but there are also problems of color distortion and incomplete defogging.To address these problems,this paper conducts research on image defogging algorithms with different requirements.The main research content of this paper is as follows:1)Aiming at the problem of atmospheric scattering model and grid artifacts,an image defogging method based on smooth dilated convolution and gating mechanism is proposed.This defogging method does not depend on the atmospheric scattering model and can degrade the image according to the input foggy blur Directly restore clear and fog-free images.Specifically,a residual block based on smooth dilated convolution is proposed,which can not only improve the task effect through smooth dilated convolution,but also improve the grid artifact problem of dilated convolution through smooth dilated convolution.At the same time,a comprehensive loss function based on mean square error loss function and perceptual loss function is proposed to train the defogging network to obtain better defogging effect.The experimental results show that the method achieves good dehazing effect on the public synthetic fog image dataset,and the restored image is clean and clear with well-preserved details.2)Aiming at the problems of insufficient utilization of convolutional layer output information and performance bottlenecks in the fusion of multi-scale feature estimation,a multi-scale dehazing network based on residual dense blocks is proposed.Specifically,refer to Grid Net,which is often used for semantic segmentation,as the main network architecture,and use the residual dense block to combine the output of the previous residual dense block and the output of each convolutional layer in the current residual dense block.Directly connected to all subsequent layers,such an operation can not only preserve the feed-forward characteristics of the network,but also extract the local dense features of the network.At the same time,a feature fusion method based on channel attention is proposed to improve the dehazing effect of the network.The simulation experiment proves that the dehazing method can more effectively remove the fog in the foggy and degraded blurred image,and the output visual effect is more consistent with the real and clear image.3)Aiming at the receptive field and attention problem of convolutional neural network,an image defogging algorithm based on self-calibrating convolution and feature attention is proposed.The algorithm introduces self-calibration convolution,without adding additional parameters,splits a standard convolution operation into multiple small convolution operations,expands the receptive field of the dehazing network,and improves the dehazing effect of the network.The channel attention and pixel attention mechanisms in the feature attention module are used to make the network pay more attention to the pixels and important channel information in the dense fog area,thereby improving the dehazing effect.It is verified by simulation experiments that this method can effectively complete the dehazing task without destroying the original information in the image,and the output haze-free image is of higher quality.