Dilated Filter Generation for CNN Feature Data Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CNN filtering processes, especially in portable terminals and embedded systems, face challenges in reducing the transfer amount of feature data and filter coefficients, and efficiently performing product-sum operations, particularly when applying dilated convolution, which complicates control and wiring due to the need to refer to feature data in preceding layers while skipping data.
Innovation Solution
A data processing apparatus and method that generates an extended filter by sequentially reading out filter coefficients and performing convolution using the extended filter, reducing the need to skip feature data and simplifying control and wiring by using a dilated filter that omits calculations for non-multiple positions of the dilation rate, thereby reducing filter coefficient transfer and processing time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If dilated convolution is performed by skipping feature data, then the reference range is extended, but control and wiring become complicated
Solution Approach 1:
Instead of skipping feature data in the spatial domain (which complicates control and wiring), the patent inverts the approach by using a dilated filter that processes all feature data sequentially. The filter coefficients are arranged such that they naturally skip positions through the dilation rate parameter, eliminating the need for complex data skipping control logic while achieving the same extended reference range effect.
Solution Approach 2:
The patent introduces a dilated filter as an intermediary structure between the feature data and the convolution operation. This filter acts as a mediator that handles the dilation function, allowing the system to maintain simple control and wiring by passing all data through the filter rather than skipping data during the convolution process.
2Area of stationary object
If filter size is increased for dilated convolution, then the reference range is extended, but filter coefficient transfer amount increases
Solution Approach 1:
The patent changes the parameter arrangement by using a dilated filter with a specific structure where filter coefficients are placed at intervals determined by the dilation rate. This parameter change allows the system to extend the reference range without transferring all possible filter coefficients, as only the necessary coefficients at dilated positions need to be transferred and processed.
3Productivity
If multiple feature data are processed in parallel, then processing speed increases, but the need to skip data complicates the architecture
Solution Approach 1:
Instead of skipping data during parallel processing (which complicates architecture), the patent inverts the approach by using a dilated filter that processes all feature data in parallel through a simplified convolution operation. The filter's dilated structure naturally handles the spacing requirement, allowing parallel processing without architectural complexity.
Data Source
AI summary
A data processing apparatus comprises a holding unit configured to hold filter coefficients of a transferred filter, a generating unit configured to generate an extended filter by extending a size of the transferred filter while sequentially reading out the filter coefficients held in the holding unit, and an arithmetic unit configured to perform convolution by using filter coefficients of the extended filter.