Image Inference Segmentation for Boundary Accuracy and Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning image processing methods struggle to achieve both accuracy and speed in inference processes, particularly at image edges or boundary parts like organ edges or tumor boundaries, due to the use of enhanced inference processes that increase processing time.
Innovation Solution
The method divides the image into a first portion where inference results are likely to be inaccurate and a second portion where they are not, performing precise enhanced inference on the first portion and faster normal inference on the second portion, then integrating the results to balance accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If enhanced inference process is used to improve accuracy at image edges and boundaries, then manufacturing precision is improved, but productivity deteriorates due to increased processing time
Solution Approach 1:
The image is divided into multiple regions including a first region containing edges and boundaries, a second region containing important parts, and a third region containing other parts. Different inference processes are applied to different regions: enhanced inference for the first region to improve accuracy at edges, and normal inference for the second and third regions to maintain processing speed. This segmentation resolves the contradiction by applying high-precision processing only where necessary.
Solution Approach 2:
Different quality levels of inference processing are applied to different parts of the image. The enhanced inference process with higher computational resources and accuracy is applied locally to edge and boundary regions, while normal inference is applied to other regions. This local differentiation allows the system to achieve high accuracy where needed without sacrificing overall processing efficiency.
2Reliability
If enhanced inference process is applied to entire image to ensure accuracy, then reliability is improved, but loss of time increases
Solution Approach 1:
The image is segmented into regions requiring high reliability (edges and boundaries in the first region) and regions where normal reliability suffices (other parts in the second and third regions). Enhanced inference is applied only to the first region to ensure reliability at critical areas, while normal inference is applied to other regions to reduce time loss.
Solution Approach 2:
Instead of applying enhanced inference to the entire image (excessive action), the system applies enhanced inference only to the extent necessary for achieving reliable results at edge and boundary regions. This partial application of enhanced processing reduces time loss while maintaining reliability where it matters most.
Data Source
AI summary
An image processing method according to an embodiment includes a specifying step, an inference step, and an integration step. In the specifying step, a first portion including a region corresponding to an anatomical site of a target and a second portion including a region different from the anatomical site are specified in the image. In the inference step, by using a deep learning model, segmentation of the region corresponding to the anatomical site is performed on the first portion and segmentation of the region different from the anatomical site is performed on the second portion, or classification and detection of an image including the region corresponding to the anatomical site is performed on the first portion and classification and detection of an image including the region different from the anatomical site is performed on the second portion. In the integration step, results of the respective processes are integrated for output.