Super-Resolution CNN Cascade Training With Network Trimming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super resolution convolutional neural networks (SRCNNs) face challenges with limited accuracy due to small structures and require excessive parameters, making real-time execution difficult, and training settings like learning rate and weight initialization are hard to tune, leading to non-convergence or local minimum issues.
Innovation Solution
Implement cascade training and trimming techniques for SRCNNs, adding intermediate layers iteratively until error thresholds are met, and reducing redundant filters to enhance efficiency and accuracy, combined with dilated convolutions to minimize computational complexity.
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 becomes prohibitively large and real-time execution becomes difficult
Solution Approach 1:
The network is divided into multiple stages with intermediate layers inserted between them. Each stage processes the output of the previous stage, allowing the network to achieve high accuracy through progressive refinement rather than requiring a single large network with all parameters at once.
Solution Approach 2:
The patent transitions from training a complete large network to training stage-by-stage with intermediate structures. This dimensional change in the training process allows manageable parameter usage at each step while achieving the accuracy of a much larger network through cascaded processing.
2Measurement precision
If the number of layers is increased to improve accuracy, then super resolution accuracy is improved, but training convergence becomes difficult and local minimum issues arise
Solution Approach 1:
The network is pre-trained in stages with intermediate structures before final refinement. This preliminary training of smaller sub-networks establishes good initial weight configurations that facilitate convergence of the complete network, preventing local minimum issues that would occur with direct training of the full large network.
Solution Approach 2:
The training process is segmented into multiple phases where intermediate layers are trained and fixed before adding subsequent layers. This segmentation of the training process ensures each sub-network converges reliably before being integrated into the larger structure.
3Measurement precision
If intermediate layers are added to improve accuracy, then super resolution accuracy is improved, but computational complexity increases
Solution Approach 1:
Intermediate layers are strategically inserted at specific positions within the network where they provide maximum benefit for accuracy improvement. Rather than uniformly increasing network depth everywhere, the patent applies intermediate structures locally at critical processing stages, optimizing the balance between accuracy gain and computational cost.
Data Source
AI summary
Apparatuses and methods of manufacturing same, systems, and methods are described. In one aspect, a method includes generating a convolutional neural network (CNN) by training a CNN having a plurality of convolutional layers, and performing cascade training on the trained CNN. The cascade training includes an iterative process of a plurality of stages, in which each stage includes inserting a residual block (ResBlock) and training the CNN with the inserted ResBlock.