Convolution Filter Noise Reduction via Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image processing techniques struggle to effectively reduce noise in image data without causing positional changes in the image, and they face challenges in generating suitable training data for prediction models, especially for optical reading applications.
Innovation Solution
A method is developed to generate an image processing filter using a neural network, which adjusts a convolution filter to reduce noise while maintaining image position integrity. This involves training the filter using a loss function that minimizes differences between output and target image data, and generating suitable training image data by setting color values for object and background pixels in a specific manner.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If aggressive noise reduction process is applied to image data, then noise is reduced, but positional changes occur in the image (edge lines shift)
Solution Approach 1:
The patent changes the parameters of the convolution filter by adjusting its weights through neural network training. The filter is trained to optimize noise reduction while maintaining positional accuracy, transforming the filter's characteristics to achieve both goals simultaneously rather than using fixed conventional filters
Solution Approach 2:
The patent implements a feedback mechanism through the loss function that evaluates both noise reduction effectiveness and positional accuracy. The loss function provides feedback during training to adjust filter weights, ensuring the filter learns to preserve edge positions while reducing noise
2Adaptability or versatility
If conventional filters are used for noise reduction, then processing is simple, but they cannot effectively reduce noise in all types of image data
Solution Approach 1:
The patent performs preliminary training of the convolution filter using training image data before actual image processing. This preliminary action creates a customized filter that adapts to specific image characteristics, making the filter more effective for particular types of image data while the training process is done in advance
Solution Approach 2:
The system uses training image data to automatically train and optimize the convolution filter without manual intervention. The filter self-adjusts its weights through the neural network training process, eliminating the need for manual filter design and adaptation for different image types
Data Source
AI summary
A method for generating an image processing filter includes: adjusting; and extracting. The adjusting inputs first training image data into a neural network to generate output image data, calculates an evaluation value based on a loss function using the output image data and second training image data, and adjusts a convolution filter so as to reduce the evaluation value. The extracting extracts data from the adjusted convolution filter as data for the image processing filter. A first training image includes noise and reproduces a test pattern. A second training image includes reduced noise and reproduces the test pattern. The loss function includes a first term and a second term. The first term specifies a magnitude of a difference between the output image data and the second training image data. The second term grows smaller as symmetry of the convolution filter relative to a filter axis of symmetry increases.