Low-Bit-Depth Neural Network Noise Estimation for Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image-quality enhancing methods using neural networks (NNs) face challenges in reducing calculation amounts while maintaining performance, particularly when quantizing weights or feature amounts into low-bit depths, which leads to degraded output accuracy.
Innovation Solution
An information processing apparatus that converts an input image of a first bit depth into a low-bit-depth image, estimates a noise component map using a neural network (NN) of a third bit depth lower than the first bit depth, and derives a noise-reduced image based on the input image and the noise component map.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the NN weight or feature amount is quantized into a low-bit depth to reduce calculation amount, then the calculation amount is reduced, but the output accuracy degrades
Solution Approach 1:
The patent segments the image processing task into two stages: first converting the input image to low-bit-depth, then performing noise component estimation. This segmentation allows the NN to operate at low bit-depth for efficiency while maintaining accuracy through the intermediate representation layer.
Solution Approach 2:
The patent introduces a low-bit-depth image representation as an intermediary between the high-bit-depth input image and the NN processing. This intermediary layer enables the NN to work with reduced precision data while still achieving accurate noise estimation through the conversion process.
2Productivity
If the NN bit depth is reduced to lower than the image bit depth, then the calculation amount is reduced, but the tones become coarse and image-quality enhancing performance lowers
Solution Approach 1:
The patent changes the bit depth parameter of the input image to match the NN's bit depth capability. By converting the input image to the same low bit-depth as the NN processes, the system eliminates the tone coarsening issue that would occur if processing high-bit-depth images through a low-bit-depth NN.
3Productivity
If simple quantization method is used to reduce bit depth, then the calculation amount is reduced, but the accuracy of the output degrades
Solution Approach 1:
The patent performs preliminary conversion of the input image to low-bit-depth before feeding it to the NN. This preliminary action prepares the data in a format suitable for the NN's bit-depth capability, ensuring accurate processing without requiring complex quantization methods during the actual inference.
Data Source
AI summary
An information processing apparatus comprises: a conversion unit configured to convert an input image of a first bit depth into a low-bit-depth image of a second bit depth lower than the first bit depth; an estimation unit configured to estimate a noise component map in the input image from the low-bit-depth image using a neural network (NN) of a third bit depth that is lower than the first bit depth and is not lower than the second bit depth; and a deriving unit configured to derive a noise-reduced image corresponding to the input image based on the input image and the noise component map.