Kernel Prediction Network for HDR Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-resolution image processing using convolutional neural networks (CNNs) is computationally intensive, and quantized image-to-image networks may not adequately represent high dynamic range (HDR) content, leading to information loss.
Innovation Solution
A kernel prediction network (KPN) is implemented with a separate data path for image signal intensity values and predicted parameters, allowing for less computationally intensive processing and high quantization, while computing coefficients for multiple kernels to be applied to different portions of an image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If convolutional neural networks are used for high-resolution image processing, then image processing quality is improved, but computational intensity increases
Solution Approach 1:
The patent segments the image processing task by dividing the image into multiple portions and computing different kernel coefficients for each portion. Instead of applying a single global kernel to the entire high-resolution image, the system processes each portion independently with locally optimized coefficients, reducing the overall computational burden while maintaining processing quality.
Solution Approach 2:
The patent implements local quality by determining that different portions of an image require different kernel coefficients optimized for their specific characteristics. Each image portion receives tailored processing with coefficients computed based on local content, rather than applying uniform processing across the entire image, thereby maintaining high quality with reduced computational requirements.
2Power
If quantized image-to-image networks are used, then computational intensity is reduced, but information loss occurs in HDR content
Solution Approach 1:
The patent changes the parameter of quantization by applying it selectively to kernel coefficients rather than to the entire image data pipeline. The kernel coefficients are quantized to reduce computational intensity, while the image signal intensity values and intermediate processing results maintain higher precision, thereby preserving HDR content information while achieving computational efficiency.
3Manufacturing precision
If multiple kernel coefficients are computed for different image portions, then processing precision is improved, but memory usage increases
Solution Approach 1:
The patent segments both the image into multiple portions and the kernel coefficients into portion-specific sets. This segmentation allows the system to store and process only the necessary coefficients for each portion independently, reducing overall memory requirements compared to storing a single large global kernel while maintaining the precision benefits of multiple specialized kernels.
Data Source
AI summary
Example methods, apparatuses, and/or articles of manufacture are disclosed that may be implemented, in whole or in part, techniques to process pixel values sampled from a multi color channel imaging device. In particular, methods and/or techniques to process pixel samples for interpolating pixel values for one or more color channels.