CNN Image Denoising with Regularization Against Overtraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Noise reduction processing using deep image prior techniques with convolutional neural networks (CNNs) achieves excellent noise reduction but suffers from image quality degradation due to overtraining, where random noise is reconstructed as the number of training iterations increases.
Innovation Solution
An image processing apparatus and method that incorporates an evaluation function with an error term and a regularization term to suppress overtraining, using a CNN for noise reduction, where the regularization term penalizes pixel value differences between adjacent pixels to stabilize the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of training iterations of the CNN is increased to improve noise reduction performance, then the noise reduction capability is enhanced, but image quality degradation occurs due to overtraining where random noise is reconstructed
Solution Approach 1:
The patent introduces a regularization term with a penalty coefficient into the loss function to change the training parameters. This regularization term penalizes large differences in pixel values between adjacent pixels, effectively controlling the overtraining phenomenon by modifying the optimization landscape without changing the fundamental DIP approach
Solution Approach 2:
The patent implements a feedback mechanism where the training process continuously evaluates the loss function comprising both the noise reduction objective and the regularization term. This feedback loop allows the system to automatically adjust the balance between noise reduction and image quality preservation by optimizing the combined loss function
Data Source
AI summary
An image processing apparatus includes a processing unit and a training unit, and performs noise reduction processing on a target image. The processing unit inputs an input image to a CNN, and outputs an output image from the CNN. The training unit uses an evaluation function, and trains the CNN based on a value of the evaluation function. The evaluation function includes an error evaluation term representing an evaluation value related to an error between the output image and the target image, and a regularization term representing an evaluation value related to a difference of pixel values between adjacent pixels in the output image. The image processing apparatus repeatedly performs respective processes of the processing unit and the training unit, and sets the output image after the respective processes are repeatedly performed a certain number of times as an image after the noise reduction processing.