CNN Feature Map Compression via Lookup Tables
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing compression methods for feature map images generated during CNN processing are not optimized for embedded systems, leading to high transmission and storage capacity requirements, and are not effectively implemented for AI model operations in devices with limited resources.
Innovation Solution
A method utilizing lookup tables and compression modes determined based on feature map characteristics to efficiently compress feature maps, reducing data size and enabling effective AI model operation in devices with limited transmission capacity by converting pixel values and predicting residuals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional JPEG, JPEG2000, PNG, or MPEG-based compression methods are applied to feature map images, then storage capacity is reduced, but the methods are not optimized for embedded systems and require high transmission capacity for effective implementation
Solution Approach 1:
The patent applies parameter changes by transforming pixel values through histogram analysis and lookup tables. The method identifies the distribution characteristics of feature map pixel values and applies specific transformation parameters (lookup tables) that map original pixel values to transformed values, optimizing the data for compression. This changes the statistical parameters of the data to achieve better compression ratios suitable for embedded systems.
Solution Approach 2:
The patent introduces lookup tables as an intermediary between the original feature map data and the compression process. These lookup tables serve as a mediator that pre-processes the pixel values by applying histogram-based transformations, making the data more amenable to compression. This intermediary step enables effective compression in embedded systems without requiring high transmission capacity.
2Quantity of substance
If conventional compression methods are used for feature maps, then storage capacity is reduced, but operation time increases due to lack of optimization for embedded systems
Solution Approach 1:
The patent applies preliminary action by performing histogram analysis and generating lookup tables before the actual compression process. The method pre-processes the feature map data by analyzing pixel value distributions and creating transformation tables in advance. This preliminary preparation optimizes the subsequent compression operation, reducing overall operation time in embedded systems while achieving effective storage reduction.
3Quantity of substance
If feature map images are stored or read in memory during CNN processing, then transmission capacity is reduced, but large transmission capacity is necessary for the image data
Solution Approach 1:
The patent transforms the statistical parameters of feature map pixel values through histogram-based lookup tables, changing the data distribution to be more compression-friendly. This parameter transformation enables more efficient memory utilization and reduces the transmission capacity required for moving data between memory and processing units during CNN operations.
Data Source
Figure 1~3
Figure 4~5
Figure 6~7A
AI summary
An electronic device and a method for controlling the same include inputting an input image into an artificial intelligence model, acquiring a feature map for the input image, converting the feature map through a lookup table corresponding to the feature map, and storing the converted feature map by compressing the feature map through a compression mode corresponding to the feature map.