Image Noise Removal Model Using DCT and DWT Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods, such as demosaicing and denoising using convolutional neural networks, face challenges in effectively removing noise and artifacts, particularly at edges, leading to artifacts like zippers or aliasing, and often require preprocessing to obtain density components, limiting the end-to-end processing capability.
Innovation Solution
A method and apparatus that utilize a noise removal model incorporating a color correlation block, discrete cosine transform (DCT) block, and discrete wavelet transform (DWT) blocks to generate feature maps highlighting high- and low-frequency components, allowing for effective noise removal and preservation of image quality by using a U-net structure with residual dense blocks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If demosaicing is performed using insufficient data from adjacent pixels, then the image reconstruction can be completed, but artifacts such as zippers or aliasing are generated at edges
Solution Approach 1:
The patent applies preliminary denoising processing to the input Bayer pattern image before demosaicing. By removing noise in advance, the subsequent demosaicing operation works with cleaner data, preventing the generation of zipper and aliasing artifacts at edges while maintaining reconstruction capability
2Reliability
If denoising is performed to remove noise from the image, then noise resistance is improved, but correct information is removed as noise leading to blur
Solution Approach 1:
The patent employs different processing strategies for different regions of the image. Edge detection is performed to identify edge regions, and different filtering operations are applied to edge regions versus non-edge regions. This local differentiation allows effective noise removal in smooth areas while preserving edge sharpness and avoiding blur
Solution Approach 2:
The patent dynamically adjusts processing parameters based on local image characteristics. By detecting edges and analyzing local variance, the system adapts the strength and type of denoising applied to different regions, maintaining noise resistance while preserving critical high-frequency information at edges
3Extent of automation
If convolutional neural network is used for demosaicing and denoising, then end-to-end processing is achieved, but the method either insufficiently removes noise or requires preprocessing that limits end-to-end capability
Solution Approach 1:
The patent integrates multiple processing functions into a unified end-to-end convolutional neural network architecture. The network combines demosaicing, denoising, and edge preservation operations in a single model that processes the Bayer pattern image directly and outputs the final RGB image, achieving both automation and effective noise removal
Data Source
AI summary
There are disclosed an apparatus and method for processing images. The apparatus for processing images according to an embodiment includes an image input unit configured to receive a first image of a Bayer pattern including noise and an image output unit configured to output a noise-removed image by removing noise from the first image using a noise removal model, and the noise removal model includes a color correlation block configured to generate a second image of the Bayer pattern including RGB correlation information about the first image from the first image by performing rearrange and convolution operations on the first image, a discrete cosine transform (DCT) block configured to generate a third image in which a high-frequency component of the first image is highlighted by applying a DCT to the first image, and one or more discrete wavelet transform (DWT) blocks configured to generate a low-frequency feature map and one or more high-frequency feature maps by applying a DWT to a first feature map generated based on the first image, the second image, and the third image, and generate a final feature map in which a high-frequency component and a low-frequency component of the first feature map are highlighted based on the low-frequency feature map and the one or more high-frequency feature maps.