Neural Network Supervisory Data Depth Conversion for Image Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks (NNs) used for image quality enhancement processing face a decrease in precision when weights and feature amounts are quantized to low-precision numerical values, particularly when the bit depth is reduced, leading to deteriorated image quality.
Innovation Solution
A method involving depth conversion processing of supervisory data to match the characteristics of target data, combined with learning to update neural network parameters, ensuring accurate image output even with reduced bit depth quantization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If weights and feature amounts of the neural network are quantized to low-precision numerical values to enable efficient calculation on equipment with limited resources, then calculation efficiency and device accessibility are improved, but precision of output data such as image resolution and gradation deteriorates
Solution Approach 1:
The patent applies preliminary action by performing depth conversion processing on supervisory data before the neural network learning process. The supervisory data is converted to match the bit depth characteristics of the target data in advance, ensuring that the learning process operates with appropriately scaled data. This preliminary preparation prevents precision loss from propagating through the learning process, allowing the network to maintain accurate gradient calculations even when working with quantized, low-precision data during inference.
2Speed
If weights are quantized to a smaller bit depth to enable high-throughput calculation instructions, then calculation speed is improved, but image quality deterioration remarkably appears
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the bit depth of supervisory data to match the characteristics of target data. Instead of using fixed high-precision supervisory data, the system converts supervisory data to appropriate bit depths (e.g., converting 16-bit supervisory data to 8-bit when processing 8-bit target data). This parameter adaptation ensures that the learning process generates weights optimized for the specific bit depth of the target application, maintaining image quality while enabling efficient low-precision inference.
Data Source
AI summary
An information processing apparatus configured to process target data by a neural network, the information processing apparatus comprising: an input data acquisition unit configured to acquire the target data; a supervisory data acquisition unit configured to acquire supervisory data; and a learning unit configured to perform learning so as to reduce an error between output data obtained by inputting, to the neural network, and processing the target data and the supervisory data, and updates a parameter of the neural network, wherein the supervisory data acquisition unit acquires the supervisory data subjected to depth conversion processing of converting a value of the supervisory data with a resolution matching a characteristic of the target data in a case where a bit depth of the supervisory data is a second bit depth smaller than a first bit depth of the target data.