Quantized Neural Network Image Processing with ISO-Based Threshold Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods using quantized machine learning models face challenges in performing proper image processing due to harmful effects caused by reduced bit precision, such as degradation in expressive power and quantization errors, especially when dealing with high ISO speeds and noisy images.
Innovation Solution
The method involves training a convolutional neural network using a quantized machine learning model, determining a threshold based on image information like ISO speed, and selecting the appropriate image information to use for processing, ensuring the model with the smallest quantization error is employed for estimation, thereby minimizing harmful effects and achieving effective image enhancement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a quantized machine learning model is used for image processing, then the model can be executed on terminals with limited memory capacity, but the expressive power degrades and quantization errors increase causing harmful effects
Solution Approach 1:
The patent applies preliminary action by determining the first threshold and selecting the appropriate machine learning model (quantized or non-quantized) before executing image processing. The determination unit evaluates image information such as ISO speed in advance and pre-selects the model that will minimize quantization errors for the specific imaging conditions, thereby preventing harmful effects before they occur.
Solution Approach 2:
The patent changes the parameter of bit precision dynamically based on image characteristics. The determination unit compares image information (e.g., ISO speed) against a threshold and selects between quantized models (lower bit precision) and non-quantized models (higher bit precision), thereby optimizing the balance between memory usage and processing accuracy for different imaging scenarios.
2Manufacturing precision
If multiple machine learning models are prepared to handle different imaging conditions, then image processing quality improves, but the memory capacity requirement increases making it impractical for terminals
Solution Approach 1:
The patent applies segmentation by dividing the machine learning models into quantized versions and non-quantized versions based on specific imaging conditions. Instead of preparing multiple complete models for all conditions, the system segments the model set into two categories that can handle different scenarios, reducing the total number of models needed while maintaining processing quality across various conditions.
Solution Approach 2:
The patent uses copying by creating quantized versions of machine learning models that approximate the functionality of full-precision models. These quantized copies consume less memory and can be used as substitutes for non-quantized models when processing images with characteristics that tolerate quantization, such as high ISO speed images where noise dominates.
3Manufacturing precision
If non-quantized machine learning models are used, then image processing accuracy is maintained, but the model size increases making it difficult to install on terminals with limited memory
Solution Approach 1:
The patent changes the bit precision parameter of the machine learning model based on image characteristics. By comparing image information (such as ISO speed) against a threshold, the system dynamically selects between quantized models (lower bit precision, smaller size) and non-quantized models (higher bit precision, larger size), thereby optimizing the balance between model size and processing accuracy for terminal devices.
Data Source
AI summary
An image processing method includes a first step of acquiring a first image and first image information about an imaging condition or a development condition corresponding to the first image, and a second step of generating a second image by enhacing the first image using a quantized machine learning model. In the second step, either the first image information or predetermined second image information is used as information to generate the second image, and a determination of whether to use either the first image information or the predetermined second image information as the information to generate the second image is based on a value relating to the first image information and a first threshold.