Radiographed Image Region Extraction with Multi-Resolution CNN Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for extracting regions in radiographed images using neural networks face challenges in achieving high speed and accuracy due to the large calculation requirements of convolutional neural networks (CNNs).
Innovation Solution
An image processing apparatus and method that includes a region extraction unit with a reduction unit, an inference unit, an enlargement unit, and a post-processing unit. This unit performs reduction processing on the input image, inference processing using a CNN, enlargement processing, and post-processing to extract regions with high accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks (CNNs) are used for region extraction processing, then accuracy is improved, but processing time increases due to large calculation requirements
Solution Approach 1:
The patent segments the region extraction process into two distinct stages: a rough extraction stage using a lightweight neural network on down-sampled images, and a precise extraction stage using CNN on the rough extraction result. This segmentation allows each stage to specialize - the first stage quickly identifies candidate regions, while the second stage refines the extraction accuracy only where needed, thus resolving the contradiction between speed and accuracy.
Solution Approach 2:
Instead of applying the computationally intensive CNN to the entire image, the patent applies it only to the rough extraction result or specific regions of interest. This partial action approach ensures that the high accuracy of CNN is utilized only where necessary, while the majority of the image is processed by the faster lightweight network, thereby reducing overall processing time while maintaining accuracy.
2Productivity
If image resolution is reduced to improve processing speed, then processing time is improved, but extraction accuracy deteriorates
Solution Approach 1:
The patent processes images at different resolutions for different purposes: down-sampled images are used for rough extraction to achieve high speed, while the original high-resolution images are used for precise extraction to ensure accuracy. This multi-resolution segmentation strategy allows the system to leverage the speed advantage of low-resolution processing while maintaining accuracy through high-resolution refinement.
Solution Approach 2:
The patent applies high-resolution processing only partially - specifically to the rough extraction result or regions of interest identified in the first stage, rather than processing the entire high-resolution image from scratch. This selective high-resolution processing maintains accuracy where it matters most while avoiding the computational cost of processing all pixels at full resolution.
Data Source
AI summary
An image processing apparatus comprising a region extraction unit configured to perform region extraction processing of extracting a region from an input image acquired based on a radiographed image, wherein the region extraction unit comprises: a reduction unit configured to perform reduction processing for the input image to generate a reduced image; an inference unit configured to perform inference processing of, using the reduced image as an input, outputting an inferred image obtained by inferring the region in the reduced image; an enlargement unit configured to perform enlargement processing for the inferred image to generate an enlarged image; and a post-processing unit configured to perform post-processing of extracting the region from the enlarged image based on a feature of a shape of the region.