| In many engineering applications,image is a major source of information,and image qualities affect the correctness of information acquisition.In addition,image qualities also affect the visual perception experience.However,the unsatisfactory imaging conditions in some scenarios result in the degradation or loss of visual information.Therefore,it is of great significance to restore clear images from degraded ones.The first two chapters of this thesis introduce the research advances on image restoration.As background knowledge,the principles of convolutional neural networks and generative adversarial networks are also briefly described.Two image restoration algorithms based on semi-supervised learning are proposed in this thesis,aiming to tackle the problems of image dehazing and underwater image enhancement,respectively.The main contributions of this thesis are summarized as follows:In chapter 3,we propose a single image dehazing algorithm based on semi-supervised domain translation and neural architecture search.Due to the domain gap between the synthetic and real-world hazy images,the dehazing algorithms trained on synthetic data show limited generalization performance on real-world hazy images.To solve this problem,we formulate dehazing as a semi-supervised domain translation problem.The proposed algorithm uses two auxiliary domain translation tasks to reduce the domain gap,leading to improved generalization performance.Dehazing and the auxiliary tasks are conducted in shared latent spaces within a unified framework,which consists of couples encoders and decoders.We also use differentiable architecture search to derive the optimal network architectures of the framework.The proposed algorithm was tested on four benchmark datasets,and its performance remarkably exceeds comparative semisupervised dehazing algorithms.Chapter 4 describes a semi-supervised underwater image enhancement algorithm based on Anderson acceleration.According to the property of the stable underwater enhancement algorithm,we designed a lightweight enhancement network operating in an iterative way.Due to the iterative structure of the network,the convergence judgment strategy is introduced to transform the network inference procedure into a fixed point problem.The Anderson acceleration algorithm is used to speed up the solution of fixed points,thereby greatly accelerating the inference speed of the neural network.In addition,a discriminator is designed to implement conditional generation adversarial learning so that the network can use both labeled data and unlabeled data for semisupervised training.The algorithm has only 0.15 M parameters,and the performance in the test set exceeds the comparison algorithms a lot.To conclude,this thesis proposes two semi-supervised algorithms to address the image restoration problems where it is infeasible to collect a large number of paired training data with ground truths.The proposed dehazing algorithm uses a domain transformation framework to improve the generalization ability of the dehazing network.In light of the convergence requirement of iterative image enhancement,the proposed underwater image enhancement algorithm increases the inference speed of the iterative image-enhancing network through Anderson acceleration.The performance of the two algorithms surpasses the comparison algorithms in their respective fields. |
