Deconvolution Kernel Weight Normalization for Image Artifact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deconvolution operations in convolutional neural networks often result in checkerboard artifacts due to varying kernel weight overlaps, leading to image quality issues and increased computational and memory requirements.
Innovation Solution
An image processing apparatus normalizes kernel weights based on their positions, dividing them into groups to ensure equal sums and applying a reliability map with smoothing functions, preventing checkerboard artifacts and adjusting image sizes effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Length of stationary object
If a deconvolution operation is performed using a deconvolution layer, then the output image size is increased, but checkerboard artifacts occur due to varying kernel overlap degrees
Solution Approach 1:
The patent applies different normalization values to different positions within the kernel based on their overlap degrees. Each kernel element is normalized according to its specific position and overlap characteristics, creating local quality variations that prevent checkerboard artifacts while maintaining overall image enlargement functionality
Solution Approach 2:
The patent modifies the normalization parameter of kernel weights based on position-dependent overlap degrees. By dynamically adjusting normalization values according to how much each kernel element overlaps with adjacent positions, the method prevents artifact formation while achieving image size increase
2Length of stationary object
If traditional deconvolution operations are performed, then image size is adjusted, but computational complexity and memory requirements increase
Solution Approach 1:
The patent computes normalization values locally for each kernel position based on overlap degree calculations, avoiding the need for complex global normalization schemes. This local computation approach reduces overall computational complexity while achieving the desired image enlargement
Data Source
AI summary
Provided is an image processing apparatus including a memory storing one or more instructions and a processor configured to execute the one or more instructions stored in the memory, wherein the processor is further configured to execute the one or more instructions to generate a second image by performing a deconvolution operation on a first image and a kernel comprising one or more weights, set values of the one or more weights based on the second image, and adjust the values of the one or more weights based on positions of the one or more weights in the kernel.