Multi-Stage Convolutional Network for Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image denoising methods, including filters and convolutional neural networks, often result in image blurring or distortion, especially when dealing with high-density noise, and are not effective in removing noise while retaining image details.
Innovation Solution
An image quality improvement method that involves preliminary denoising filtering followed by a multi-stage convolutional network model with multiple sub-models of different architectures, which are sequentially connected and trained using a multi-stage loss function to generate an optimization image that effectively removes noise and restores image details.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If linear filters or non-linear filters are used for denoising, then noise removal is achieved, but image edge blur or image distortion occurs
Solution Approach 1:
The patent segments the denoising process into multiple stages with different filter types (e.g., median filter, bilateral filter, Gaussian filter) applied sequentially or selectively to different regions. This segmentation allows each filter to target specific noise characteristics while preserving image edges through adaptive filter selection and parameter adjustment.
2Object-affected harmful factors
If the filter mask size is increased to improve high-density noise removal, then denoising effect improves, but image blur increases
Solution Approach 1:
The patent implements dynamic filter mask size adjustment based on local noise density and image content characteristics. The mask size adapts automatically - larger masks are applied to high-density noise regions while smaller masks preserve details in edge regions. This dynamic adaptation is achieved through algorithms that analyze local variance, edge detection, and noise density estimation to adjust filter parameters in real-time.
3Manufacturing precision
If convolutional neural network models are used for denoising, then image quality is improved, but the models are not effective in removing high-density noise
Solution Approach 1:
The patent creates a composite denoising system that combines traditional filtering methods with convolutional neural networks. The system integrates multiple processing stages including preprocessing filters, CNN-based denoising modules, and post-processing refinement. This composite approach leverages the strength of traditional filters in handling high-density noise while utilizing CNNs for overall image quality enhancement and detail preservation.
4Object-affected harmful factors
If multiple convolutional neural network models are trained for different noise densities, then denoising effectiveness is improved, but training complexity and cost increase
Solution Approach 1:
The patent employs a single convolutional neural network model with adaptive parameter adjustment instead of training multiple models for different noise densities. The network dynamically modifies its internal parameters (filter weights, activation thresholds, pooling sizes) based on the input image's noise characteristics. This parameter adaptation is achieved through techniques like batch normalization with adaptive moments, learnable filter sizes, and noise-level estimation modules that adjust processing intensity without requiring separate model training.
Data Source
AI summary
An image quality improvement method and an image processing apparatus using the same are provided. Denoising filtering is performed to an original image by a filter to obtain a preliminary processing image. The preliminary processing image is input to a multi-stage convolutional network model to generate an optimization image through the multi-stage convolutional network model. The multi-stage convolutional network model includes multiple convolutional network sub-models, and these convolutional network sub-models respectively correspond to different network architectures.