Neural Network Image Processing for Stable Inference Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image processing techniques using deep learning, such as neural networks, face issues where image tint changes after white balance correction or High Dynamic Range (HDR) processing, leading to decreased inference accuracy and noise retention in low-luminance portions.
Innovation Solution
An image processing apparatus that acquires and processes training and correct answer images to update neural network parameters, minimizing errors after white balance correction and HDR processing, ensuring stable inference accuracy and reduced tint changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If white balance correction or HDR processing is applied to the RAW image inferred by the neural network, then the image is adjusted for viewing, but the tint of the image changes and inference accuracy decreases
Solution Approach 1:
The patent applies the same white balance correction or HDR processing to both the training image and the correct answer image before inputting them to the neural network. This preliminary processing ensures that the neural network learns the relationship between processed images, so that when the same processing is applied to the inferred image during actual use, the tint remains stable and inference accuracy is maintained.
2Device complexity
If the neural network is trained without applying processing to training images, then the learning process is simpler, but the inferred image shows tint changes after processing
Solution Approach 1:
The patent applies the same white balance correction or HDR processing to both the training image and the correct answer image before inputting them to the neural network. This preliminary processing ensures that the neural network learns the relationship between processed images, so that when the same processing is applied to the inferred image during actual use, the tint remains stable and inference accuracy is maintained.
Data Source
AI summary
An apparatus includes one or more processors that function as an image acquisition unit configured to acquire a training image and a correct answer image, a generation unit configured to input the training image to a neural network to generate an output image, an error acquisition unit configured to subject each of the correct answer image and the output image to processing for adjusting a color signal value, and acquire an error between the correct answer image and the output image that have been subjected to the processing, and an update unit configured to update parameters of the neural network based on the acquired error.