Convolutional Neural Network Decoder for Image Morphing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image morphing technologies using convolutional neural networks are inefficient due to the complexity and high parameter count of conventional architectures, leading to slower processing speeds and longer processing times.
Innovation Solution
A computer-implemented method using a convolutional neural network with a feature extraction sub-network, a morpher, and a decoder sub-network, where the decoder sub-network employs densely connected computational modules with convolutional blocks and un-pooling blocks to efficiently morph input images with style features, reducing the number of parameters and enhancing processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional convolutional neural network architectures are used for image morphing, then image quality can be maintained, but processing speed decreases and computational complexity increases
Solution Approach 1:
The decoder sub-network is segmented into multiple densely connected computational modules, where each module contains multiple convolutional blocks. This segmentation allows for more efficient feature propagation and reduces computational redundancy compared to conventional sequential architectures.
Solution Approach 2:
The patent merges feature maps from previous convolutional blocks within densely connected computational modules, allowing multiple feature maps to be combined and processed together. This merging approach reduces the overall number of parameters and computational operations while maintaining feature representation quality.
2Productivity
If conventional convolutional neural network architectures are used for image morphing, then comprehensive feature extraction can be achieved, but the number of parameters increases leading to slower processing
Solution Approach 1:
Densely connected computational modules merge feature maps from multiple previous convolutional blocks, reducing the total number of parameters needed while maintaining comprehensive feature extraction. This is achieved by reusing features across multiple blocks rather than processing them independently.
Solution Approach 2:
The architecture dynamically adjusts feature flow through densely connected modules where feature maps are propagated and reused across multiple convolutional blocks. This dynamic feature reuse reduces parameter count while maintaining processing efficiency.
3Productivity
If densely connected computational modules with multiple convolutional blocks are used in the decoder sub-network, then processing speed and efficiency improve, but architectural complexity increases
Solution Approach 1:
The decoder is segmented into standardized densely connected computational modules, each containing a fixed number of convolutional blocks. This segmentation creates a modular architecture that improves processing efficiency through systematic feature propagation while managing complexity through repetition of standardized units.
Solution Approach 2:
The densely connected computational modules serve multiple functions: they perform feature extraction, feature fusion, and feature propagation simultaneously within a unified structure. This multi-functionality improves processing efficiency without proportionally increasing architectural complexity.
Data Source
AI summary
A computer-implemented method using a convolutional neural network is provided. The computer-implemented method includes processing an input image through the convolutional neural network to generate an output image including content features of the input image morphed with style features of a style image. The convolutional neural network includes a feature extraction sub-network, a morpher, and a decoder sub-network. Processing the input image through convolutional neural network includes extracting style features of the style image to generate a plurality of style feature maps using the feature extraction sub-network; extracting content features of the input image to generate a plurality of content feature maps using the feature extraction sub-network; morphing the plurality of content feature maps respectively with the plurality of style feature maps to generate a plurality of output feature maps using the morpher; and reconstructing the plurality of output feature maps through the decoder sub-network to generate the output image.