Image Processing With Feature-Map Convolution Skipping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods to increase throughput in convolutional neural network (CNN) processing face challenges as increasing the size of small regions for calculation speedup results in fewer zero-value regions, leading to insufficient calculation speedup and difficulty in changing the size of small regions.
Innovation Solution
An image processing device and method that performs convolution processing for each small region of an input feature map, skipping processing for regions with identical or sequentially similar features, using a threshold to determine if the features are identical or similar within a predetermined threshold, and utilizing a preliminary calculation result table to output processed results directly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the size of small regions is increased to increase throughput, then calculation speedup is insufficient because fewer zero-value regions are found
Solution Approach 1:
The input feature map is divided into multiple small regions, allowing parallel processing of different regions. This segmentation enables the system to process multiple areas simultaneously, increasing throughput while maintaining the ability to skip zero-value regions for calculation speedup.
Solution Approach 2:
The system performs preliminary determination of whether each small region contains zero values or identical features before executing convolution processing. This preliminary check allows the system to skip unnecessary calculations for regions that would produce zero or identical results, reducing overall calculation time while maintaining throughput.
2Productivity
If the size of small regions is increased, then fewer cases have all zero values, reducing calculation speedup
Solution Approach 1:
The system applies different processing strategies to different small regions based on their local characteristics. Regions with zero values or identical features are handled differently (skipped or processed once) compared to regions with diverse values, optimizing calculation efficiency for each local area while maintaining overall throughput.
Solution Approach 2:
The system changes the processing parameters based on the content of each small region. By detecting whether regions have zero values, identical features, or diverse values, the system adjusts its processing approach accordingly, enabling efficient calculation while maintaining throughput even with larger region sizes.
3Productivity
If the size of small regions is increased, then it becomes difficult to change the size of small regions
Solution Approach 1:
The system dynamically determines the appropriate small region size based on the input feature map characteristics and processing requirements. This dynamic adjustment allows the system to adapt to different scenarios, maintaining flexibility in region size selection while optimizing throughput for each specific case.
Data Source
AI summary
When convolution processing is performed, an input feature map to be an input of the convolution processing is divided into small regions, and in a case where features constituting the small region correspond to a predetermined feature or a feature of a small region processed in past, the convolution processing is not performed for the small region, and a result of processing for the predetermined feature or a result of processing in past is output as a processing result for the small region.