Deep Learning Based Image Restoration with the Approach of Noise Classification
| Autor: | WU, YI-KAI, 吳翊楷 |
|---|---|
| Rok vydání: | 2017 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 105 In recent years, deep learning has been successfully implemented in many different purposes because of the huge improvements in computer hardware and algorithms. Different neural networks are designed to fit different purposes, such as go (game), medicine, and image recognition. This thesis proposes a method which combines convolution neural networks and classic image restoration method to classify images before the restoration. Most of the testing data of restoration method are produced from known original images by adding noises on each pixel. However, there are no referenceable labels for restoration method in practices. The NCM (Noise Classification Model) proposed in the thesis can be used in predicting the noise of the input images, so the restoration method can restore the images with the prediction. The NCM model is designed to classify the deviation values of images produced by AWGN (Adaptive White Gaussian Noise). The deviation values of the prediction will be the parameter of NLM (Non-Local Algorithm) which is one of famous restoration methods in in the field of restoration. The model is trained by tens of thousands of small images which are all produced by AWGN based on different random single colors. After training, the NCM is tested with different noised Lena pictures. Although the results of predictions are not very accurate, the curve of the deviation values of the prediction and input images is a monotonic increasing function. That means the NCM will not predict two different noised images into the same deviation value. Finally, the proposed NCM is simply a new possible way to restore images, despite there are still some other improved or expand ideas. This study proposes a method which predicts the input images before the restoration and that solves the problem not having a referenceable variable in practice. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|