Withdraw
Loading…
Towards practical deep learning based image restoration model
Zhou, Yuqian
Loading…
Permalink
https://hdl.handle.net/2142/113899
Description
- Title
- Towards practical deep learning based image restoration model
- Author(s)
- Zhou, Yuqian
- Issue Date
- 2021-12-02
- Director of Research (if dissertation) or Advisor (if thesis)
- Hasegawa-Johnson, Mark
- Doctoral Committee Chair(s)
- Hasegawa-Johnson, Mark
- Committee Member(s)
- Liang, Zhi-Pei
- Bhat, Suma
- Shi, Humphrey
- Department of Study
- Electrical & Computer Eng
- Discipline
- Electrical & Computer Engr
- Degree Granting Institution
- University of Illinois at Urbana-Champaign
- Degree Name
- Ph.D.
- Degree Level
- Dissertation
- Keyword(s)
- Image Restoration
- Deep Learning
- Image Inpainting
- Image Denoising
- Abstract
- Image Restoration (IR) is a task of reconstructing the latent image from its degraded observations. It has become an important research area in computer vision and image processing, and has wide applications in the imaging industry. Conventional methods apply inverse filtering or optimization-based approaches to restore images corrupted in ideal cases. The limited restoration performance on ill-posed problems and the low-efficient iterative optimization processes prevents such algorithms from being deployed to more complicated industry applications. Recently, the advanced deep Convolutional Neural Networks (CNNs) begin to model the image restoration as learning and inferring the posterior probability in a regression model, and successfully achieved remarkable performance. However, due to the data-driven nature, the models trained with simple synthetic paired data (e.g, bicubic interpolation or Gaussian noises) cannot be well adapted to more complicated inputs from real data domains. Besides, acquiring real paired data for training such models is also very challenging. In this dissertation, we discuss the data manipulation and model adaptability of the deep learning based image restoration tasks. Specifically, we study improving the model adaptability by understanding the domain difference between its training data and its expected testing data. We argue that the cause of image degradation can be various due to multiple imaging and transmission pipelines. Though complicated to analyze, for some specific imaging problems, we can still improve the performance of deep restoration models on unseen testing data by resolving the data domain differences implied in the image acquisition and formation pipeline. Our analysis focuses on digital image denoising, image restoration from more complicated degradation types beyond denoising and multi-image inpainting. For all tasks, the proposed training or adaptation strategies, based on the physical principle of the degradation formation or based on geometric assumption of the image, achieve a reasonable improvement on the restoration performance. For image denoising, we discuss the influence of the Bayer pattern of the Camera Filter Array (CFA) and the image demosaicing process on the adaptability of the deep denoising models. Specifically, for the task of denoising RAW sensor observations, we find that unifying and augmenting the data Bayer pattern during training and testing is an efficient strategy to make the well-trained denoising model Bayer-invariant. Additionally, for the RGB image denoising, demosaicing the noisy RAW images with Bayer patterns will result in the spatial-correlation of pixel noises. Therefore, we propose the pixel-shuffle down-sampling approach to break down this spatial correlation, and make the Gaussian-trained denoiser more adaptive to real RGB noisy images. Beyond denoising, we explain a more complicated degradation process involving diffraction when there are some occlusions on the imaging lens. One example is a novel imaging model called Under-Display Camera (UDC). From the perspective of optical analysis, we study the physics-based imaging processing method by deriving the forward model of the degradation, and synthesize the paired data for both conventional and deep denoising pipeline. Experiments demonstrate the effectiveness of the forward model and the deep restoration model trained with synthetic data achieves visually similar performance to the one trained with real paired images. Last, we further discuss reference-based image inpainting to restore the missing regions in the target image by reusing contents from the source image. Due to the color and spatial misalignment between the two images, we first initialize the warping by using multi-homography registration, and then propose a content-preserving Color and Spatial Transformer (CST) to refine the misalignment and color difference. We designed the CST to be scale-robust, so it mitigates the warping problems when the model is applied to testing images with different resolution. We synthesize realistic data while training the CST, and it suggests the inpainting pipeline achieves a more robust restoration performance with the proposed CST.
- Graduation Semester
- 2021-12
- Type of Resource
- Thesis
- Permalink
- http://hdl.handle.net/2142/113899
- Copyright and License Information
- Copyright 2021 Yuqian Zhou
Owning Collections
Graduate Dissertations and Theses at Illinois PRIMARY
Graduate Theses and Dissertations at IllinoisManage Files
Loading…
Edit Collection Membership
Loading…
Edit Metadata
Loading…
Edit Properties
Loading…
Embargoes
Loading…