Research On Multi-scale Connected Image Dehazing Algorithm Based On Deep Learning

In recent years,with the rapid development of computer vision,technologies such as license plate recognition and target tracking have been widely used in people’s lives.However,in the foggy environment,the quality of the image captured by the multimedia imaging device is relatively serious,which has a certain impact on the visual processing task.Therefore,how to effectively dehaze in the captured image is a key step in post-vision tasks.In this paper,the specific reasons for image degradation in haze environment are studied in detail,and the deep learning image dehazing algorithm is studied.The main research contents of this paper are as follows:1.Study the specific process and principle of the image of the target scene captured by the imaging device in the haze environment,The basic composition and application of convolutional neural network are introduced,and the classic dehazing algorithm of Dahaze Net based on deep learning is introduced,experimental simulation is carried out and the defects in the algorithm are analyzed.2.A new deep learning single image dehazing algorithm is proposed to solve the large error in the parameter estimation of the traditional dehazing algorithm.First,a multi-scale convolutional neural network is used to learn the transmittance of the input haze image;then,the extracted transmittance is refined by using the guided filter to keep the edge smooth to obtain a transmittance that is closer to the input image;Then use the dark channel prior knowledge to estimate the global atmospheric light value;Finally,based on the atmospheric scattering model,the image dehazing is restored.The experimental comparison and analysis with a variety of classic dehazing algorithms show that the algorithm proposed in this paper has a significant improvement in dehazing effect and indicators.3.Aiming at the problem of the loss of detailed information during the training of the deep learning network and the incomplete dehazing when the fog concentration is unevenly distributed,a region-adaptive dehazing algorithm based on multiple connections is proposed.The algorithm network framework is similar to the UNet3+network,which is mainly composed of an encoding part,an upsampling layer,a downsampling layer and a decoding part.A novel interactive multi-scale connected is used between the encoding network and the decoding network to provide sufficient feature information during training.Using multiple rectangular pixel blocks,the haze area in the haze map is divided into multiple rectangular areas with "approximately" uniform haze concentration distribution to extract rich background information.In order to avoid the loss of feature information in the sampling process,discrete wavelet transform is used for scale change operation.Compared with the newer deep learning dehazing algorithm,the peak signal-to-noise ratio(PSNR)is increased by 1.77 d B,the structural similarity(SSIM)is increased by 0.02,and the dehazing effect is significantly improved.