Research On Image Dehazing Technology Based On Deep Learning

Due to the light absorption and scattering of atmospheric particles,the light reflected by the objects decreases in hazy weather.And influenced by various ambient light,hazy weather can lead to blurred imaging results as well as the problem of insufficient image brightness and contrast.This not only affects the viewability of images,but also poses a great challenge to image processing tasks that require high image quality such as military and unmanned vehicles.Therefore,the research of image dehazing is of great significance to various image processing systems and people’s production life.Traditional dehazing algorithms are prone to distortion,unnaturalness,poor detail recovery,and low generalizability.Deep learning image dehazing algorithms suffer from haze residue,lack of contrast and saturation of the original image,and lack of s tability of the output high-resolution image.The main research contents are as follows.(1)The relevant algorithms for image dehazing are analyzed,and deep learning techniques such as atmospheric scattering theory,convolutional neural networks,residual neural networks,generative adversarial networks,and image quality evaluation indexes are studied(2)An improved High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs(Pix2PixHD)dehazing algorithm Pix2 Pix HD-SOS based on the Strengthen-Operate-Subtract(SOS)enhancement module is designed.firstly,the haze map input to the generator is dehazed,and then the dehazed image and its corresponding haze-free image are input to the discriminator for discrimination.By adding the SOS enhancement module to the decoder of the generation network,the dehazing effect of the algorithm is improved in terms of detail feature recovery.Experimental results on NYU Depth,I-haze,RESIDE,and O-haze indoor and outdoor image simulations and real datasets show that the method has better results in terms of subjective visual as well as objective evaluation metrics for the dehazed images compared with algorithms such as DCP,AHE,and MSBDN-DFF.(3)Aiming at the problems of insufficient image contrast a nd insufficient detail recovery in the Pix2PixHD-SOS algorithm,an improved Pix2 Pix HD-SOS image haze removal algorithm Pix2 Pix HD-SOS-CBAM based on the attention mechanism is proposed.By introducing CBAM attention mechanism in the residual block between encoder and decoder,the haze feature-rich part is highlighted to suppress other unnecessary features.The network is optimized by the inclusion of structural similarity(SSIM)loss and the increase of feature matching loss weights.Experiments show that this algorithm can generate high-resolution haze removal images with higher contrast and richer details.