Tile-Based Convolutional Neural Network Processing with Overlapping Pixel Exclusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems using convolutional neural networks (CNNs) face challenges in achieving favorable processing results due to the division of images into tiles, which leads to incomplete convolutional computation at the edges of tiles, resulting in reduced processing accuracy and increased processing time.
Innovation Solution
The proposed solution involves an image processing apparatus that executes convolutional computation processing by obtaining tiles corresponding to partial regions of an image and controlling the computation unit to exclude overlapping pixels from the target of the convolutional computation processing, thereby optimizing the processing time and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the image is divided into tiles for processing, then the buffer size limitation is addressed and processing becomes feasible, but the processing time increases due to overlapping pixel regions
Solution Approach 1:
The patent extracts and excludes overlapping pixels from the convolutional computation processing. By identifying pixels that appear in multiple tile regions and excluding them from redundant computation, the system maintains the tile-based processing approach while eliminating the time penalty associated with processing overlapping regions multiple times.
2Measurement precision
If tiles are configured to include overlapping pixel regions, then favorable processing results are achieved at edge portions, but the processing time period increases
Solution Approach 1:
The patent removes overlapping pixels from the computation target after they have been used to achieve accurate edge processing. The overlapping pixels are included in tile configurations to ensure favorable processing results at boundaries, but then extracted from subsequent computation to eliminate redundant processing time.
Solution Approach 2:
The system performs preliminary convolutional computation on overlapping pixels when they are first encountered in a tile, then excludes them from computation in subsequent tiles. This preliminary action ensures that the computation is performed only once despite the pixels appearing in multiple tile regions.
3Measurement precision
If convolutional computation is performed on the entire image without division, then processing accuracy is maintained, but the buffer size requirement exceeds the CNN computation unit capacity
Solution Approach 1:
The patent divides the image into multiple tiles that can be processed within the buffer capacity of the CNN computation unit. Each tile is processed independently with convolutional computation, allowing the system to handle large images that would otherwise exceed buffer limitations while maintaining processing feasibility.
Data Source
AI summary
An image processing apparatus executes convolutional computation processing in a neural network with respect to input data, obtains a plurality of tiles that respectively correspond to partial regions in an image and performs control so as to cause the convolutional computation processing to be executed while using each of the plurality of tiles as the input data. The control is performed so that, with respect to at least a part of the plurality of tiles, overlapping pixels which are included in the at least the part of the plurality of tiles and which correspond to the same region in the image as another tile are excluded from a target of the convolutional computation processing.