Image Restoration Based On Dynamic Neural Networks

During acquiring and processing images,they usually degrade to low-quality images(e.g.,hazy,blurry,or noisy images)by factors such as haze and rain.These degraded images influence not only human perceptual but also object detection and tracking algorithms.Nowadays,deep learning methods,represented by convolution neural networks,have obtained numerous researches and applications in single restoration tasks,such as dehazing,deraining,and denoising.However,they are difficult to recover real-world images,which are usually influenced by multiple factors.For solving these issues,this work proposes dynamic neural networks designed for low-quality images,which are infected by single,multiple or real-world factors.Specifically,this work achieves the following tasks:1.By considering degraded appearance of low-quality images have different patterns,this work first proposes a recurrent feedback network based on image content.This network can dynamically restore different local regions by the recurrent feedback mechanism.Moreover,given different kinds of low-quality images,the difficulty of learning clear appearance is usually different.Therefore,this work further explores how the appearance generation methods influence the final performance,and empirically proposes a fusion block of appearance generation methods,leading to achieve multiple restoration tasks in a single framework.2.Because there are variant degradation in images infected by multiple factors,algorithms designed for removing certain degradation tend to perform poor.Therefore,this work first explores performances of the above algorithms,and then find that they always suffer from learning features of other degradation.Based on this observation,this work proposes two dynamic fusion networks for utilizing learning abilities of different networks,thus improving the restoration performance.3.In the above algorithms,these networks are trained by synthetic images,suffering from the image number and the domain discrepancy between synthetic and real-world images.Thus,this work proposes the embedding images for jointly using real-world and synthetic images to train deep networks.Furthermore,this work introduces the ranking loss function for enhancing the generalization ability of deep neural networks to real-world images.According to literature review,the above methods are the first semi-supervised approach achieved by data augmentation.