An Information-theoretic analysis of generative adversarial networks for image restoration in physics-based vision

  • Xudong KANG

Student thesis: MPhil Thesis (Lingnan)


Image restoration in physics-based vision (such as image denoising, dehazing, and deraining) are fundamental tasks in computer vision that attach great significance to the processing of visual data as well as subsequent applications in different fields. Existing methods mainly focus on exploring the physical properties and mechanisms of the imaging process, and tend to use a deconstructive idea in describing how the visual degradations (like noise, haze, and rain) are integrated with the background scenes. This idea, however, relies heavily on manually engineered features and handcrafted composition models, which can be theories only in ideal conditions or hypothetical models that may involve human bias or fail in simulating true situations in actual practices. With the progress of representation learning, generative methods, especially generative adversarial networks (GANs), are considered a more promising solution for image restoration tasks. It directly learns the restorations as end-to-end generation processes using large amounts of data without understanding their physical mechanisms, and it also allows completing missing details damaged information by involving external knowledge and generating plausible results with intelligent-level interpretation and semantics-level understanding of the input images. Nevertheless, existing studies that try to apply GAN models to image restoration tasks dose not achieve satisfactory performances compared with the traditional deconstructive methods. And there is scarcely any study or theory to explain how deep generative models work in relevant tasks.

In this study, we analyzed the learning dynamics of different deep generative models based on the information bottleneck principle and propose an information-theoretic framework to explain the generative methods for image restoration tasks. In which, we study the information flow in the image restoration models and point out three sources of information involved in generating the restoration results: (i) high-level information extracted by the encoder network, (ii) low-level information from the source inputs that retained, or pass directed through the skip connections, and, (iii) external information introduced by the learned parameters of the decoder network during the generation process.

Based on this theory, we pointed out that conventional GAN models may not be directly applicable to the tasks of image restoration, and we identify three key issues leading to their performance gaps in the image restoration tasks: (i) over-invested abstraction processes, (ii) inherent details loss, and (iii) imbalance optimization with vanishing gradient. We formulate these problems with corresponding theoretical analyses and provide empirical evidence to verify our hypotheses and prove the existence of these problems respectively.

To address these problems, we then proposed solutions and suggestions including optimizing network structure, enhancing details extraction and accumulation with network modules, as well as replacing measures of training objectives, to improve the performances of GAN models on the image restoration tasks. Ultimately, we verify our solutions on bench-marking datasets and achieve significant improvement on the baseline models.
Date of Award21 Jul 2022
Original languageEnglish
Awarding Institution
  • Lingnan University
SupervisorHaoran XIE (Supervisor) & Man Leung WONG (Co-supervisor)

Cite this