Research In Image Dehazing Model Based On Deep Convolutional Network

With the popularity of intelligence and informatization,images play an increasingly important role in our life as the medium that most intuitively reflects real-time information,but the frequent occurrence of foggy weather in practical scenarios seriously affects the clarity of images captured by intelligent devices,thus hindering the effective performance of subsequent advanced vision tasks.Therefore,people have started to continuously work on image clarifying techniques to improve the quality of blurred images.This article is based on the convolutional neural network structure,and analyzes and improves the existing image dehazing methods at home and abroad,which have insufficient accuracy in restoring hazy image texture details,network information overload,low efficiency of single image dehazing,lack of effective use of real foggy images,and insufficient model robustness,The main work is as follows:(1)A Balance Information Fusion Network(BIFN)is proposed to reconstruct clear images.BIFN is a fusion of three tasks representing different levels of background,target,and semantic features of the target image.Each task utilizes a balanced convolutional perception module and a dense residual module for global information perception and feature enhancement extraction,respectively.The feature fusion modules of each layer cross combine high and low layer features to achieve information exchange.In addition,our model training uses L1 smoothing loss and perceptual loss as the joint loss function in the supervision strategy to ensure that the restored image is close to the original clear image in visual sense.The experiment proves that our proposed balanced information fusion network greatly improves the quality of blurred images and improves the efficiency of single image dehazing.(2)A semi supervised double branch network(SBDNet)is proposed that combines supervised and unsupervised training methods.The supervision part uses the balanced information fusion network BIFN mentioned in the previous chapter to constrain model training by constructing positive and negative image samples through comparative learning,in order to enhance the ability of model generalization.In the unsupervised branch based on generative adversarial networks,real fogged images are introduced into model training.The discriminator uses a Markov discriminator,while the generator uses a dense residual attention network based on U-Net architecture.Through continuous confrontation between the discriminator and the generator,the generative network approximates the distribution pattern of real images.The experiment has proven that SDBNet has good defogging performance on both synthetic and real-world datasets.(3)We conducted a field investigation on X Airport and preprocessed the foggy images collected by intelligent vision equipment to construct an airport foggy image sample,aiming to improve the operational safety of the airport through image dehazing technology.Finally,the two image defogging models proposed in the previous section were applied to the airport dataset for experimental exploration.The results showed that SDBNet had better defogging effects than BIFN in airport foggy environments,but the single image defogging efficiency was lower.