Research On Image Deraining Method Based On Deep Neural Network

With the development of the computer era,weather image restoration techniques have been widely used and become an important task in the field of computer vision.Adverse outdoor weather(e.g.,rain,snow)will lead to degradation of image quality,which not only severely affects the visual quality of images,but also has a significant impact on various visual tasks(e.g.,object tracking,object detection,and image segmentation).Traditional methods using dictionary learning and sparse representation techniques can only extract superficial information from images,making it almost impossible for them to fully recover complete structural texture and edge knowledge.In contrast,deep learning-based methods bring significant improvements in feature representation,allowing them to extract raindrop morphology and background structural texture edge information.However,these methods do not fully explore the blur morphology of raindrops,are prone to produce artifacts that are inconsistent with the surroundings,and such methods require large-scale manual annotation of datasets.This thesis addresses the limitations of current image de-raindrop methods and proposes three image de-raindrop methods based on deep neural network.1.Raindrop removal method for single image based on a selective generation adversarial network.The generator network extracts local raindrop morphological features via a selective skip connection module and obtains global structural features of the image via a self-attentive module.The generator uses a combination of global and local information to generate a more global view of the image with finer details.The discriminator evaluates the validity of the output to further improve the quality of the output image.2.Selective generative adversarial network for raindrop removal from a single image.This method further explores the morphological characteristics of raindrops,proposes a new raindrop image modeling method,and uses a multi-scale framework to fuse the cross-scale internal correlation information of raindrops.Due to the accurate learning of the raindrop blur level and the collaborative representation learning of raindrops across scales,we predict the raindrop distribution more correctly and output visually more convincing image effects.3.Unsupervised image raindrop removal method based on detangling generative adversarial network.The method utilizes a detangling mechanism to separate the raindrop features in an image and then self-supervises the image de-raindrop effect via the Cycle GAN method.The fact that no manual labeling of data is required makes the method more practical and more in line with the development of modern computer vision tasks.Finally,in order to demonstrate the effectiveness of the three methods proposed in this paper in the image de-raindrop task,they are validated on several classical datasets.Qualitative and quantitative experimental results show that the three methods proposed in this paper have a superior performance in generating clearer,more coherent and visually more plausible reconstruction results than previous models.