Shared Convolution Neural Network for High-Resolution Image Illumination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies using multiple neural networks for different functions suffer from increased training time, processing time, and resource usage due to overlapping functional fields, and struggle with high resolution image processing, leading to inefficient memory occupancy and network loading times.
Innovation Solution
A neural network design that utilizes a common convolution field for feature extraction, with separate fully connected layers for filter selection and parameter optimization, allowing for efficient image quality enhancement by resizing low illumination intensity high resolution images to low resolution, applying a filter scheme, and converting them back to high illumination intensity high resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple neural networks with separate convolution fields are used for different image processing functions, then each function can be performed independently, but training time increases
Solution Approach 1:
The patent merges multiple separate convolution fields into a single shared convolution field that serves multiple neural networks. This shared convolution field extracts features once and makes them available to multiple functions, eliminating redundant processing while maintaining functional independence through separate fully connected layers for each specific task.
Solution Approach 2:
The shared convolution field is designed to perform multiple functions simultaneously by extracting features that can be utilized by different neural networks for different image processing tasks. This universal feature extraction capability allows one component to serve multiple purposes, reducing overall system complexity and training time.
2Adaptability or versatility
If multiple neural networks with separate convolution fields are used for different image processing functions, then each function can be performed independently, but processing time increases
Solution Approach 1:
The patent merges multiple separate convolution fields into a single shared convolution field that serves multiple neural networks. This shared convolution field extracts features once and makes them available to multiple functions, eliminating redundant processing while maintaining functional independence through separate fully connected layers for each specific task.
3Adaptability or versatility
If multiple neural networks with separate convolution fields are used for different image processing functions, then each function can be performed independently, but memory occupancy increases
Solution Approach 1:
The patent merges multiple separate convolution fields into a single shared convolution field that serves multiple neural networks. This shared convolution field extracts features once and makes them available to multiple functions, eliminating redundant processing while maintaining functional independence through separate fully connected layers for each specific task.
4Manufacturing precision
If image processing is applied directly to high resolution images, then image quality is maintained, but processing time and resources increase excessively
Solution Approach 1:
The patent segments the image processing task into two stages: first processing a downsampled low-resolution version of the image to determine filter parameters, then applying those parameters to the full high-resolution image. This segmentation allows the computationally intensive neural network operations to be performed on smaller data, while still achieving high-quality results on the original high-resolution image.
Solution Approach 2:
The patent performs preliminary processing on a downsampled version of the image to determine optimal filter parameters before applying filters to the full high-resolution image. This preliminary action on reduced-size data reduces computational burden while preserving the ability to produce high-quality output on the original image.
Data Source
AI summary
A method for enhancing illumination intensity of an image according to an embodiment of the present disclosure may include receiving an image, determining a filter application scheme by using a neural network model trained to enhance the illumination intensity of the image, and outputting an image having enhanced illumination intensity by applying, to the image, the determined filter application scheme. The neural network model trained using machine learning includes a first group of layers for extracting feature information of the image, a second group of layers for determining a type of filter based on the extracted feature information, and a third group of layers for determining the optimal parameter for the filter based on the extracted feature information. The neural network model may be provided through an external server in an IoT environment using a 5G network.