CNN Feature Map Compression Using Repeated Small Regions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in increasing throughput in convolutional neural network (CNN) operations is that larger small regions in the input feature map lead to fewer instances where all values are zero, hindering calculation speedup, and larger data sizes result in longer memory access times, severely impacting calculation speed.
Innovation Solution
An image processing device that divides output feature maps into small regions and skips convolution processing for regions with identical or sequentially identical features, using a preliminary calculation result table to output processed results directly, thereby reducing data size and memory access time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image data is compressed to reduce data amount, then transmission efficiency improves, but image quality deteriorates due to lossy compression
Solution Approach 1:
The patent divides image processing into multiple stages: initial compression for efficient transmission, then separate processing of luminance and chrominance components at the receiving end. This segmentation allows different compression ratios for different color components, maintaining overall image quality while achieving transmission efficiency.
Solution Approach 2:
The patent applies different compression strategies to different parts of the image data. Specifically, luminance components are preserved with higher quality while chrominance components use more aggressive compression, as human vision is more sensitive to brightness variations than color variations. This local quality approach resolves the contradiction between compression efficiency and image quality.
2Manufacturing precision
If multiple image signals are processed separately, then color accuracy improves, but processing time increases
Solution Approach 1:
The patent merges the processing of luminance and chrominance signals into a unified processing framework. By combining the processing steps for multiple color components into a single integrated system, the patent achieves color accuracy comparable to separate processing while significantly reducing overall processing time through parallel operations and optimized data flow.
3Manufacturing precision
If image data is transmitted in original quality, then image quality is preserved, but data transmission volume increases
Solution Approach 1:
The patent changes the representation parameters of image data by separating luminance and chrominance components and applying different compression ratios to each. This parameter transformation allows the system to reduce overall data transmission volume while maintaining perceived image quality, as the human visual system is more tolerant of chrominance compression than luminance compression.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
When an output feature map to be an output of convolution processing is output, the output feature map is output to a storage unit for each of divided small regions. When each small region is output to the storage unit, in a case where a feature included in the small region is the same as a predetermined feature or a feature of a small region output in the past, the predetermined feature or the feature of a small region output in the past is compressed and output to the storage unit.