Neural Network Image Restoration via Frequency Channel Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image compression technologies, such as JPEG, result in blurry edges and lost high-frequency details, making it difficult to restore high-quality images from compressed images, and existing restoration methods using CNNs often produce blurry outputs due to the loss of high-frequency information.
Innovation Solution
A method involving a trained artificial neural network that processes images by creating a feature map, an estimated class probability distribution map, and concatenating it with a frequency map to produce a clearer output image, using classification techniques to restore lost high-frequency data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image compression technology (JPEG) is used to reduce data size, then network load is reduced and transmission efficiency is improved, but high-frequency details are lost and image quality deteriorates
Solution Approach 1:
The patent segments the image processing task into multiple specialized neural network components: an encoder for feature extraction, a frequency processing module for high-frequency restoration, and a decoder for image reconstruction. This segmentation allows each component to focus on specific aspects of image restoration, thereby recovering high-frequency details that were lost during compression while maintaining transmission efficiency
Solution Approach 2:
The patent introduces a frequency processing module as an intermediary between the encoder and decoder. This module specifically processes high-frequency components that were removed during JPEG compression, acting as a mediator to restore lost information before final image reconstruction, thus solving the contradiction between compression efficiency and detail preservation
2Manufacturing precision
If CNN-based restoration methods are used to recover lost information, then some image quality improvement is achieved, but the output images appear blurry due to ineffective high-frequency restoration
Solution Approach 1:
The patent applies local quality by dedicating specific neural network layers and processing modules to handle different frequency components of the image. The frequency processing module specifically targets high-frequency regions with specialized operations, ensuring that these critical but previously lost components are restored with appropriate detail and sharpness, rather than applying uniform processing across the entire image
3Measurement precision
If loss minimization functions (MSE, MAE) are used in CNN training, then the mapping function between input and output images is optimized, but edges and textures where high-frequency loss occurred are difficult to restore effectively
Solution Approach 1:
The patent changes the training approach by using a two-stage training process with different loss functions for different stages. The first stage uses standard loss functions for general feature learning, while the second stage employs perceptual loss and gradient-based loss functions that specifically reward sharp edges and fine textures. This parameter change in the training objective enables effective restoration of edges and textures that standard pixel-wise loss functions cannot achieve
Data Source
AI summary
A method of processing an image for improving image quality is provided. The method includes inputting an image into a first artificial neural network as an input variable to create a feature map and inputting an image into a second artificial neural network as an input variable to create an estimated class probability distribution map for each of frequency channels. Further, the method may comprise determining a class for each of the frequency channels on the basis of a probability distribution included in the estimated class probability distribution map, creating an estimated class map that includes a determined class, converting the estimated class map into an estimated frequency map, combining the estimated frequency map with the feature map to create a combined feature map, and inputting the combined feature map into a third artificial neural network as an input variable to create an output image.