Neural Network Black Floating Adjustment for Image Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise reduction techniques using deep neural networks struggle with individual sensor differences and variations in black floating amounts between training and inference times, limiting their effectiveness in reducing noise in images.
Innovation Solution
A neural network-based image processing system that adjusts black floating levels in input images to match those during training, using a convolutional neural network (CNN) for noise reduction, and includes a cloud server for generating training data and performing noise reduction inference on edge devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If a deep neural network is used for noise reduction, then noise is reduced, but black floating varies due to sensor differences and temperature changes
Solution Approach 1:
The patent applies preliminary action by adjusting the black floating level of the input image before performing noise reduction inference. The adjustment unit modifies the input image to match the black floating characteristics of the training data, ensuring the neural network operates under conditions similar to its training environment. This pre-adjustment step resolves the contradiction by preparing the image in advance to maintain consistent black floating behavior across different sensors and temperature conditions.
Solution Approach 2:
The patent employs parameter changes by dynamically adjusting the black floating level parameter of the input image based on the difference between the actual input image characteristics and the training data characteristics. The adjustment unit calculates the required parameter change and applies it to the input image before inference, thereby adapting the input to match the expected training conditions and maintaining reliable black floating suppression across varying sensor conditions.
2Measurement precision
If analog gain and digital gain are applied to increase sensitivity, then image sensitivity is improved, but noise amount increases
Solution Approach 1:
The patent converts the harmful effect of increased noise (resulting from analog and digital gains) into a benefit by using the neural network to specifically target and remove noise while preserving the enhanced sensitivity. The noise reduction unit leverages the trained model to distinguish and eliminate noise components from the high-sensitivity image signal, thereby transforming the noise problem into an opportunity for selective noise removal that maintains the sensitivity improvements.
3Manufacturing precision
If black floating adjustment is performed to match training conditions, then noise reduction effectiveness is improved, but processing complexity increases
Solution Approach 1:
The patent reduces processing complexity by performing the black floating adjustment as a preliminary step before the computationally intensive noise reduction inference. The adjustment unit prepares the input image in advance by matching its black floating characteristics to the training data, so that the subsequent neural network inference can proceed more efficiently with inputs that already match expected conditions, thereby improving effectiveness without proportionally increasing overall complexity.
Data Source
AI summary
An information processing apparatus configured to reduce noise in an image using a trained neural network includes at least one processor and at least one memory having instructions stored thereon that, when executed by the at least one processor, cause the at least one processor and the at least one memory to cooperate to adjust a black floating of an input image so as to be closer to a black floating of an image used at a time of training of the neural network, and perform inference processing on a noise-reduced image of the adjusted image, using the neural network trained to suppress the black floating.