Convolutional Neural Network Image Processing via Sub-Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for accelerating convolutional neural networks in computing-constrained environments, such as mobile devices, are ineffective for pixel-level tasks like image segmentation and super-resolution due to significant performance degradation from weight reduction techniques.
Innovation Solution
The method involves converting high-resolution input images or feature maps into smaller target images or feature maps, which are processed by convolutional neural network blocks, and then enlarging the resolution of the output feature maps to maintain information integrity and improve computational speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If weight reduction techniques are applied to accelerate convolutional neural networks, then processing speed is improved, but model performance significantly degrades
Solution Approach 1:
The input image is divided into multiple sub-images, each processed independently by the neural network. This segmentation allows parallel processing of smaller data units, improving throughput while maintaining the full resolution information across all sub-images, thus avoiding performance degradation from weight reduction.
Solution Approach 2:
The patent transforms the problem from processing a single high-resolution image to processing multiple low-resolution sub-images. By changing the dimensional organization of input data (from one large image to multiple smaller images), the system achieves faster processing while preserving overall image quality through aggregation of results.
2Productivity
If the resolution of input images is reduced to decrease calculation load, then processing speed is improved, but information loss occurs
Solution Approach 1:
Instead of reducing the resolution of a single image (which causes information loss), the patent segments the original high-resolution image into multiple sub-images. Each sub-image is processed at lower resolution, but the collective set of sub-images preserves all original image information, avoiding the information loss problem.
Solution Approach 2:
The patent applies resolution reduction partially to only certain regions (sub-images) rather than the entire image. This partial application of downscaling maintains calculation speed benefits while preserving overall information integrity through the combination of multiple processed sub-images.
Data Source
AI summary
An image processing method comprises obtaining an input image; converting the input image or a feature map of the input image into a plurality of target input images or target feature maps, wherein a resolution of each of the target input images or the target feature maps is smaller than a resolution of the feature map of the input image or the input image, and pixels at the same position in each of the target input images or the target feature maps are of a neighborhood relationship with the input image or the feature map of the input image; processing at least a part of the plurality of target input images or target feature maps by one or more convolution blocks in a convolutional neural network; and increasing a resolution of a feature map output from the one or more convolution blocks in the convolutional neural network.