Feature Up-sampling Networks for CNN Computation Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional CNNs face challenges in reducing computation time while maintaining high accuracy for feature map generation, as the amount of computations is proportional to the size of the input image, leading to decreased detection accuracy when image size is reduced.
Innovation Solution
The implementation of Feature Up-sampling Networks (FUN) in a CNN, which includes a down-sampling block to reduce image size, followed by convolution operations and up-sampling blocks to rescale feature maps, allowing for the generation of feature maps with reduced computation time and preserved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If the size of the input image is reduced to decrease computation time, then the computation time is reduced, but the detection accuracy deteriorates
Solution Approach 1:
The network is segmented into multiple stages with different processing resolutions. The input image is divided into multiple scales, and each scale is processed separately through filter blocks and up-sampling blocks, allowing computations to be performed on smaller segments while preserving overall accuracy through multi-scale feature fusion
Solution Approach 2:
The patent introduces a new dimension of feature representation by generating feature maps at multiple scales (different resolutions) and fusing them together. This multi-scale dimensional approach allows the system to capture both fine details and broader contextual information, maintaining detection accuracy while enabling efficient computation at reduced resolutions
2Productivity
If the size of the input image is reduced to reduce the number of channels, then the computation time is reduced, but the performance of feature map generation deteriorates
Solution Approach 1:
The down-sampling block performs preliminary reduction of the input image size before the feature maps are generated by the filter blocks. This preliminary action reduces the computational burden on subsequent processing stages while the up-sampling blocks restore the feature map resolutions, ensuring that feature generation performance is maintained despite the initial size reduction
Solution Approach 2:
The up-sampling blocks act as intermediaries between the down-sampled feature maps and the final output. These blocks restore the spatial resolution of feature maps that were processed at lower resolutions, mediating between computational efficiency requirements and the need for high-quality feature representations for accurate detection
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A learning method for improving performance of a CNN by using Feature Up-sampling Networks is disclosed. The learning method includes steps of: (a) allowing the down-sampling block to acquire a down-sampling image; (b) allowing each of a (1-1)-th to a (1-k)-th filter blocks to respectively acquire each of a (1-1)-th to a (1-k)-th feature maps; (c) allowing a specific up-sampling block to (i) receive a particular feature map from its corresponding filter block, and (ii) receive another specific feature map from its previous up-sampling block, and then rescale a size of the specific feature map to be identical to that of the particular feature map and (iii) apply certain operations to the particular feature map and the rescaled specific feature map to generate a feature map of the specific up-sampling block; and (d)(i) allowing an application block to acquire an application-specific output and (ii) performing a first backpropagation process.