Image Classification Heat Maps for Fracture Surface Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fracture surface analysis devices primarily focus on the accuracy of estimation results without considering the intermediate results, making it difficult to visually confirm which region of the fracture surface image is being focused on during estimation.
Innovation Solution
An image classification system that generates and displays section image data with frames indicating their positions on the specimen image, using a trained neural network to estimate classification elements, and superimposes display information such as classification elements, probabilities, heat maps, and statistical data on the specimen image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a trained neural network is used to estimate fracture mode from fracture surface image, then the estimation accuracy is improved, but it becomes difficult to visually confirm which region of the image is being focused on during estimation
Solution Approach 1:
The fracture surface image is divided into multiple section images, and the neural network performs estimation for each section independently. This segmentation allows the system to process complex images in manageable parts while maintaining estimation accuracy, and enables visual tracking of which specific regions are being analyzed.
Solution Approach 2:
Section image frames are introduced as intermediary visual elements that bridge the gap between the input fracture surface image and the final estimation result. These frames display the position and classification of each section, serving as a visual mediator that shows users which regions are being focused on during the estimation process.
2Device complexity
If only the estimation result is shown to the user, then the system complexity is reduced, but the user cannot determine the validity of the estimation
Solution Approach 1:
The system provides feedback to users by displaying section image frames with classification results and heat maps that show which regions contributed most to the estimation. This feedback mechanism allows users to verify the validity of estimations without significantly increasing system complexity, as the feedback is generated automatically from the neural network's intermediate outputs.
Solution Approach 2:
Instead of showing raw neural network internal states, the system creates simplified visual copies in the form of section image frames and heat maps that represent the estimation process. These visual copies make the estimation validity checkable while keeping the underlying system relatively simple.
3Loss of information
If section image data with display information is generated and superimposed on the specimen image, then visual confirmation of attention regions is improved, but the processing complexity increases
Solution Approach 1:
The system merges multiple pieces of information (section image frames, classification elements, heat maps, and statistical data) into a single composite display that is superimposed on the original specimen image. This merging approach consolidates the processing requirements while providing comprehensive visual feedback, rather than requiring separate processing for each type of information.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
PROBLEM TO BE SOLVED: To provide an image classification system with which a region of attention, etc., of an image can be easily confirmed visually, and the validity of an estimation can be determined. SOLUTION: An image classification system comprises a data receiving part which receives specimen image data, a data generating part which generates a plurality of section image data contained in said specimen image data as a partial image data and having smaller size than said specimen image data, a data estimation part which estimates, for each of the plurality of section image data, to which classification element each of the plurality of section image data is classified among a plurality of classification elements set in advance corresponding to a feature of the surface, by using a neural network trained in advance by teacher image data, and a display part that generates display information related to the section image data or the classification element, and superimpose and displays the generated display information on a specimen display image that is a display image of the specimen image data.