Cascade Trained Super Resolution Network for Real-Time Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution convolutional neural networks (SRCNNs) face limitations in accuracy and efficiency due to their small structure and high parameter count, leading to slow convergence and increased computational complexity, making real-time execution challenging.
Innovation Solution
The proposed solution involves a cascade trained super resolution convolutional neural network (CT-SRCNN) approach, which starts with a small network and gradually adds layers, using cascade training and network trimming to reduce complexity, along with dilated convolutions to enhance efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the size of SRCNN is increased to improve accuracy, then super-resolution accuracy is improved, but the number of parameters increases prohibitively and execution speed decreases
Solution Approach 1:
The network is divided into multiple stages with progressively added layers. Each stage trains a subset of layers independently before merging, allowing the large network to be trained in manageable segments that can be executed efficiently
Solution Approach 2:
Layers are pre-trained independently in separate stages before being merged into the complete network. This preliminary training of individual components allows the final network to achieve high accuracy without requiring prohibitively large parameter counts during execution
2Measurement precision
If the size of SRCNN is increased to improve accuracy, then super-resolution accuracy is improved, but device complexity increases
Solution Approach 1:
The complex network is segmented into multiple manageable stages, each with a specific number of layers. This segmentation reduces the perceived complexity during training and deployment while maintaining the capability for high accuracy in the final merged network
Solution Approach 2:
The network structure is dynamically adjusted during training by adding layers in stages. The complexity increases progressively rather than all at once, allowing for manageable training processes and flexible optimization at each stage
3Productivity
If the number of parameters is reduced to improve execution speed, then real-time processing is enabled, but training convergence becomes difficult
Solution Approach 1:
Each layer subset is pre-trained independently to convergence before being merged with other subsets. This preliminary convergence of individual components ensures that when combined, the complete network maintains reliable convergence properties despite having fewer total parameters
Solution Approach 2:
The training process uses feedback from each stage's convergence behavior to adjust subsequent training parameters. This feedback mechanism ensures reliable convergence even as the network size is optimized for faster execution
Data Source
AI summary
A method and apparatus are provided. The method includes generating a dataset for real-world super resolution (SR), training a first generative adversarial network (GAN), training a second GAN, and fusing an output of the first GAN and an output of the second GAN.