Denoised Image Detail-Loss Detection with CNN Error Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for denoising ray-traced images often result in the loss of fine details, particularly in areas with high variance or far from the camera, leading to blurred textures and structures, which current metrics like mean absolute error and structural similarity index measure (SSIM) fail to adequately detect.
Innovation Solution
A computer-implemented method forms a dataset for training a Convolutional Neural Network (CNN) architecture by computing differences between normalized features of denoised and reference images, generating an error map to identify areas of degradation, using a pre-trained VGG-19 architecture, and adding these pairs to the dataset for supervised learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If denoising algorithms are applied to reduce noise in ray-traced images, then noise variance is reduced, but fine details and textures are lost leading to blurred images
Solution Approach 1:
The patent applies different processing treatments to different regions of the image based on their local characteristics. Regions with fine details (like leather textures) are identified and preserved with minimal denoising, while regions with coarse noise are denoised more aggressively. This is achieved through analyzing local variance and detail metrics to adaptively control the denoising strength in each region.
Solution Approach 2:
The denoising process is made dynamic by iteratively adjusting the denoising parameters based on intermediate results. The algorithm performs multiple passes, each time adapting the denoising strength based on the current state of the image, allowing it to preserve details that would be lost in a single static denoising pass.
2Device complexity
If standard metrics like mean absolute error or SSIM are used to evaluate denoised images, then computational complexity is low, but detection of detail loss is inadequate
Solution Approach 1:
The evaluation metric is segmented into multiple components that analyze different aspects of image quality separately. Instead of using a single holistic metric, the patent divides the analysis into local variance measurement, detail preservation assessment, and noise reduction evaluation, allowing each component to specialize in detecting specific types of quality issues.
Solution Approach 2:
The patent introduces a new dimension to image quality evaluation by analyzing the frequency domain characteristics and local structural patterns beyond traditional pixel-level comparisons. This multi-dimensional approach enables detection of detail loss that single-metric methods miss.
Data Source
AI summary
A computer-implemented method for forming a dataset configured for learning a Convolutional Neural Network (CNN) architecture including an image feature extractor. It comprises providing pairs of images, each pair comprising a reference image and a respective denoised image. For each pair of images, the method provides the pair of images to a pre-trained CNN architecture similar to the one the formed dataset will be configured for. The method computes an error map representing a difference between a first normalized feature of the denoised image and a second normalized feature of the reference image, the first and second normalized features being the output of a same layer of the pre-trained CNN architecture and adds the respective denoised image and the error map to the dataset. This constitutes an improved solution with respect to forming a dataset for learning a CNN architecture to identify areas of degradation generated by a denoiser.