Image Processing Apparatus Non-Adjacent Pixel Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks (CNNs) face inefficiencies in processing time due to the need for continuous memory addresses and the inability to skip feature data during dilated convolution operations, which affects processing efficiency, especially in resource-constrained devices like portable terminals and embedded systems.
Innovation Solution
An image processing apparatus and method that allows for the skipping of non-adjacent pixels in feature images during convolution operations, utilizing a first obtaining unit to retrieve pixels and a calculating unit to perform convolution based on these obtained pixels, enabling efficient data transfer and processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If continuous memory addresses are used for convolution operations, then memory access efficiency is improved, but dilated convolution operations cannot skip feature data which reduces processing efficiency
Solution Approach 1:
The patent segments the feature image into multiple blocks and processes them independently. Each block can be processed with its own memory access pattern, allowing non-sequential access to different blocks while maintaining efficiency within each block. This segmentation enables dilated convolution to skip feature data without requiring continuous memory addresses across the entire feature image.
Solution Approach 2:
The patent transforms the one-dimensional memory access sequence into a two-dimensional block structure. By organizing feature data into blocks with spatial coordinates, the system can access non-adjacent pixels efficiently by jumping between blocks rather than sequentially through linear memory addresses. This dimensional transformation enables efficient random access patterns required for dilated convolution.
2Measurement precision
If all feature data is processed in convolution operations, then complete feature extraction is achieved, but unnecessary calculations increase processing time
Solution Approach 1:
The patent extracts and processes only the necessary feature data points for dilated convolution operations. By identifying which pixels are actually needed based on the dilation rate and convolution kernel, the system skips unnecessary feature data points. This selective extraction maintains complete feature extraction for relevant positions while eliminating wasted calculations on positions that will be skipped.
Solution Approach 2:
The patent applies partial action by processing only the subset of feature data that is actually needed for the convolution operation. Instead of uniformly processing all feature data, the system performs calculations only where the convolution kernel overlaps with non-zero coefficients, reducing total operations while maintaining extraction precision for the required features.
3Productivity
If non-adjacent pixels are referenced during convolution, then processing efficiency is improved, but memory access patterns become non-sequential which may reduce efficiency
Solution Approach 1:
The patent performs preliminary actions by pre-organizing feature data into blocks and pre-calculating the positions of pixels that will be accessed during dilated convolution. By preparing the data structure in advance with known access patterns, the system can efficiently jump to required positions without costly runtime memory search or random access optimization. This preliminary organization mitigates the speed penalty of non-sequential access.
Solution Approach 2:
The patent implements dynamic memory access patterns that adapt to the specific dilation rate and convolution kernel size. The memory access strategy changes based on the operation parameters, optimizing the balance between skipping feature data and maintaining access speed. This dynamic approach allows the system to switch between different access patterns (e.g., block-based vs. stride-based) depending on the computational requirements.
Data Source
AI summary
An image processing apparatus comprises a first obtaining unit configured to obtain a pixel from a feature image, and a calculating unit configured to perform a convolution operation based on a pixel obtained by the first obtaining unit. The first obtaining unit is capable of obtaining non-adjacent pixels from the feature image.


