Spatially Variant Kernels for Edge-Preserving Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods using deep neural networks apply the same kernel to all pixels, leading to deteriorated performance due to varying pixel characteristics, especially in image denoising tasks.
Innovation Solution
An image processing apparatus that utilizes a convolutional neural network to generate adaptive image processing by calculating similarity information between each pixel and its adjacent pixels, generating a spatially variant kernel based on regional characteristics, and applying different kernels to maintain detailed edges and textures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the same kernel is applied to all pixels in image processing, then the processing is simple and fast, but the performance deteriorates due to varying pixel characteristics
Solution Approach 1:
The patent applies different kernels to different pixels based on their local characteristics. Specifically, it calculates similarity information between each pixel and its adjacent pixels, then selects or generates appropriate kernels based on regional characteristics, ensuring that each pixel is processed with a kernel suited to its local features rather than using a uniform kernel across the entire image.
Solution Approach 2:
The patent dynamically adapts the kernel selection or generation process based on the input image characteristics. By calculating similarity information and determining regional characteristics in real-time, the system adjusts which kernels are applied where, making the processing adaptive rather than static.
2Manufacturing precision
If different kernels are applied according to pixel characteristics, then image processing performance is improved, but the device complexity increases
Solution Approach 1:
The patent segments the image processing task by dividing it into regions based on pixel similarity characteristics. It calculates similarity information between adjacent pixels and groups them into regions with comparable characteristics, then applies appropriate kernels to each region. This segmentation approach manages complexity by handling similar pixels together rather than individually.
Solution Approach 2:
The patent performs preliminary calculation of similarity information between pixels and their adjacent neighbors before the actual kernel application. This pre-processing step identifies regional characteristics in advance, allowing the system to prepare and select appropriate kernels beforehand, which streamlines the subsequent processing and reduces overall system complexity.
3Manufacturing precision
If adaptive kernel generation is performed based on regional characteristics, then detailed edge processing and texture are maintained during denoising, but the processing time increases
Solution Approach 1:
The patent applies adaptive kernel processing selectively based on regional characteristics rather than uniformly to all pixels. By identifying regions with specific characteristics (such as edges or textures) and applying appropriate kernels only where needed, the system achieves effective denoising with detail preservation while avoiding unnecessary complex processing in uniform regions.
Solution Approach 2:
The patent performs preliminary calculation of similarity information and regional characteristic determination before kernel application. This advance preparation allows the system to identify which regions require detailed processing and which can use simpler approaches, optimizing the balance between detail preservation and processing time.
Data Source
AI summary
An image processing apparatus, including a memory configured to store one or more instructions; and a processor configured to execute the one or more instructions stored in the memory to obtain similarity information indicating a similarity between each pixel of a plurality of pixels included in a first image and an adjacent pixel of the each pixel; generate a weight map including weight information corresponding to the each pixel, based on the similarity information; generate a spatially variant kernel including a plurality of kernels corresponding to the plurality of pixels, based on the weight map and a spatial kernel including weight information based on a location relationship between the each pixel and the adjacent pixel; and generate a second image by applying the spatially variant kernel to the first image.