Image Feature Extraction Using Karatsuba Convolution Submaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolution neural networks face high computational resource consumption and inefficient feature extraction processes, particularly in operations like face feature extraction and image processing, due to existing optimization methods like FFT and Winograd algorithms that either increase memory bandwidth or require excessive adder resource consumption.
Innovation Solution
An image feature extraction method and apparatus that utilizes a Karatsuba algorithm for convolution operations, involving preprocessing, zero-padding, and overlapping feature submaps to optimize resource consumption, specifically reducing multiplier and adder resource usage through efficient multiplication and addition operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If FFT algorithm is used to optimize convolution operation, then calculation speed is improved, but memory bandwidth consumption increases
Solution Approach 1:
The patent divides the feature map into multiple submaps and processes them separately through cutting and zero-padding operations. This segmentation allows the convolution operation to be performed on smaller data blocks, reducing the memory bandwidth requirements compared to processing the entire feature map at once with FFT, while still achieving efficient calculation through the Karatsuba algorithm.
Solution Approach 2:
The patent transforms the convolution operation parameters by applying zero-padding to submaps and using Karatsuba algorithm which changes the mathematical approach from standard convolution to a more efficient polynomial multiplication-based method. This parameter transformation reduces computational complexity and memory bandwidth consumption while maintaining calculation accuracy.
2Loss of energy
If Winograd algorithm is used to optimize convolution operation, then computational resource consumption is reduced, but adder resource consumption increases excessively
Solution Approach 1:
The patent segments the feature map into submaps and processes them independently. This segmentation reduces the overall computational burden and adder resource requirements compared to applying Winograd algorithm to the entire feature map, as each submap requires fewer adders while the total resource consumption is optimized through the Karatsuba approach.
Solution Approach 2:
The patent replaces the traditional convolution mechanism with the Karatsuba algorithm, which substitutes the standard convolution computation with a polynomial multiplication-based approach. This substitution reduces both multiplier and adder resource consumption compared to Winograd algorithm, achieving lower overall computational resource usage while maintaining efficiency.
3Measurement precision
If standard convolution operation is performed on large feature maps, then feature extraction accuracy is maintained, but computational resource consumption is high
Solution Approach 1:
The patent divides the large feature map into multiple smaller submaps through cutting operations. This segmentation maintains feature extraction accuracy by preserving the spatial relationships and boundary information through zero-padding, while significantly reducing computational resource consumption by processing smaller data blocks that require fewer computational resources.
Solution Approach 2:
The patent performs preliminary zero-padding on each submap before convolution operation. This preliminary action ensures that boundary features are preserved and accuracy is maintained, while the preprocessing step enables more efficient subsequent convolution operations that consume fewer computational resources compared to processing the original large feature map directly.
Data Source
AI summary
An image feature extraction method and apparatus are provided, along with a computer device and a readable storage medium. After obtaining the target convolutional kernel of the feature extraction convolutional neural network, this application will cut the input feature map obtained from the preprocessed image according to the preset image size into multiple abutting input feature submaps. Each input feature submap will be zero-padded to obtain the preprocessed feature submap. Then, for each preprocessed feature submap, the Karatsuba algorithm will be used to reduce the multiplier resource consumption during the convolution operation between the target convolutional kernel and the preprocessed feature submap, effectively suppressing the growth of adder resource consumption during the convolution operation, resulting in the output feature submap. Subsequently, multiple output feature submaps will be overlapped and the boundary of the overlapped feature map will be cropped to obtain the output feature map.