Convolutional Neural Network Architecture for Multiscale Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional CNNs lack a network structure that is compatible with the use of multiscale representation information and multilayer network learning, making accurate image processing difficult.
Innovation Solution
An image processing apparatus that includes a main unit using a first convolutional neural network with at least two layers to convert an input image into a first feature amount, a second convolutional neural network to convert the input image into a second feature amount, and a third convolutional neural network to combine these feature amounts and generate an output image, enabling the use of multiscale expression information and multilayer network learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional CNN structures are used, then the network can be implemented with standard architectures, but the network cannot effectively utilize multiscale representation information and multilayer network learning simultaneously
Solution Approach 1:
The network is segmented into multiple independent CNN modules (first CNN, second CNN, third CNN) that can be configured with different architectures and parameters. Each module processes features at different scales independently, allowing the system to utilize multiscale representation information while maintaining multilayer network learning capabilities through the modular structure.
2Reliability
If a single CNN architecture is used, then the implementation is simple, but the network cannot achieve high accuracy by combining multiscale information and multilayer learning
Solution Approach 1:
Multiple CNN modules with different configurations are merged into a unified network architecture. The first CNN extracts features at one scale, the second CNN extracts features at another scale, and the third CNN combines these features through addition operations. This merging enables the network to achieve high accuracy by integrating multiscale information while maintaining a relatively simple overall structure through the use of standard convolutional operations.
3Measurement precision
If conventional CNN structures are used, then the network can process images, but the processing accuracy is insufficient due to inability to utilize multiscale expression information
Solution Approach 1:
The network extends the feature extraction process into a multiscale dimension by configuring different CNN modules to operate at different scales. The first CNN and second CNN process the same input image at different scales, and the third CNN combines these multiscale features. This dimensional extension in scale space enables the network to achieve higher feature extraction accuracy by capturing patterns at multiple resolutions simultaneously.
Data Source
AI summary
An image processing apparatus includes a main unit configured to convert an input image into information of a first feature amount using a first convolutional neural network having at least two layers, an input unit configured to convert the input image into information of a second feature amount using a second convolutional neural network, and an output unit configured to convert information of a third feature amount generated by adding the information of the first feature amount and the information of the second feature amount to each other, into an output image using a third convolutional neural network.