Sequential CNNs for Steel Microstructure Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The interpretation of microscopic images of steel samples for microstructure analysis is subjective and dependent on the expertise of metallographers, leading to variability and lack of objectivity in classification results.
Innovation Solution
A method utilizing two sequentially trained convolutional neural networks (CNNs) to classify steel microstructures, where the first CNN identifies structural phases like ferrite, martensite, and others, and the second CNN specifically differentiates between bainitic and non-bainitic ferrite, enabling more reliable and reproducible analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If classification is performed manually by metallographers based on empirical knowledge, then flexibility and adaptability are maintained, but objectivity and reproducibility deteriorate due to subjective interpretation
Solution Approach 1:
The patent replaces the manual mechanical classification process performed by metallographers with an automated image processing system using convolutional neural networks. The system processes electron micrographs through multiple CNNs that automatically classify microstructure phases, eliminating subjective human interpretation while maintaining adaptability through programmable classification rules and configurable parameters for different steel types and phases.
2Productivity
If a single comprehensive classification system is used, then all microstructure phases can be classified in one step, but classification precision deteriorates for specific difficult-to-distinguish phases like bainitic ferrite
Solution Approach 1:
The patent divides the classification task into multiple sequential stages using different CNNs. The first CNN performs broad classification of major phases (ferrite, martensite, austenite, pearlite, carbide), while the second CNN specifically focuses on differentiating bainitic ferrite from non-bainitic ferrite. This segmented approach allows each network to specialize in specific phases, improving precision for difficult distinctions while maintaining overall productivity through automated processing.
3Adaptability or versatility
If comprehensive training data including all microstructure phases is provided to one CNN, then all phases can be classified simultaneously, but training complexity and computational requirements increase
Solution Approach 1:
The patent segments the training process into two separate CNN training tasks. The first CNN is trained on comprehensive data covering all major phases, while the second CNN is trained specifically on ferrite phase differentiation. This reduces the complexity of each individual training task, making the system more manageable while maintaining comprehensive classification capability through the combined output of both networks.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for analyzing microstructures in steel samples, wherein a steel sample is processed by separation and abrasive treatment as well as etching and an electron microscopic image (1) of the surface is produced, which is fed into a first trained Convolutional Neural Network (CNN) (10a), wherein the training data of the first CNN include electron microscopic images, among others, with associated classification data relating to image regions of microstructure phases of "ferrite".After the image (1) is classified by the first CNN, it is subsequently fed into a second trained CNN (10b). The training data of the second CNN includes images with associated classification data for image regions, at least from the microstructure phases "Bainitic Ferrite" (13a) and "Non-Bainitic Ferrite" (13b), as training classes. The classification by the second CNN (10b) is restricted, based on the classification by the first CNN, to image regions (13) that were classified as "Ferrite" by the first CNN (10a). The combined classification of both the first (10a) and the second (10b) CNN is output, with the classification of the second CNN being output instead of the classification of the first CNN for the regions classified by the second CNN.