Speaker
Description
Grain density is an important microstructural property in metallographic analysis to determine the steel mechanical properties. In a previous work [3], we have shown the benefit of using deep neural networks to automatically estimate grain densities. That work focused primarily on austenitic materials to establish the deep learning architecture and applied a transfer-learning concept to also allow for grain density analysis in martensitic alloys. Compared to austenitic data, martensitic data have a higher degree of variability (e.g. extreme contrast, lack of edges defining grain boundaries, etc.) and look more like textures than well-structured grain arrangements. In order to improve the performance of the deep neural networks, we propose here the use of image pre-processing methods to enhance the input data, suppressing unwanted distortions and emphasizing relevant visual patterns.
Texture-based descriptors have long been used for object classification. Among a number of approaches, Local Binary Patterns (LBP)[4] have become popular due to their simplicity and good performance. LBP compute a local representation of textures by comparing each pixel to the pixel’s surroundings. One of the most important properties of LBPs are their robustness to monotonic grayscale changes (e.g. illumination variations and contrast). One of the parameters of the LBP is the size of the local neighborhood around each pixel. That parameter allows us to compute the LBPs at different scales: the smaller the scale, the finer a texture is described, and the larger, the coarser the description. For improving the classification of martensitic data, we exploit the same Convolutional Neural Network (CNN) structure (based on the well-established ResNet architecture [2]) previously used but instead of training the network with raw image input data, we provided as input their texture-based (LBP) representation.
As we cover grain density classes from 2.5 to 12, thus exhibiting a high variety in scales, a single texture descriptor cannot cover the broad range of texture scales. To account for that, we trained four similar networks at different LBP scales ( 1, 3, 5 and 9 pixels radii). The classification of an image by a single CNN results in a normalized distribution (over all grain classes). To combine the resulting normalized distributions of the four different CNNs, a Support Vector Machine (SVM) classifier [1] is used to obtain a final (combined) normalized distribution.
For martensitic alloys, we show that instead of using raw image data as input but a texture-based (LBP) representation of that same data at four different scales, the combined classification result outperforms the result from our previous work. On real training, validation and test sets of respectively 643, 215 and 215 images (as basis for data augmentation) spread over 11 grain classes, the CNN with texture based preprocessed inputs reached a Top1 accuracy of 86.4% with a Mean Absolute Error (MAE) of 0.16. In comparison, the best performing single CNN using raw image input data reached a Top1 accuracy of 83.3% with a Mean Absolute Error (MAE) of 0.46. While the increase in accuracy amounts to only 3%, the decrease in MAE is pretty important: roughly a half grain size class. That speaks for the pertinence of using LBP as input for classifying martensitic data. It also shows how dedicated preprocessing can help in reducing the amount of required annotated ground truth data, which is extremely difficult to obtain for the high variety of martensitic appearances.
References:
[1] Cortes, C. and Vapnik, V., "Support-vector networks", in Machine Learning 20 (3), 1995, pp. 273–297.
[2] He, K. et al, “Deep residual learning for image recognition,” in Conference on Computer Vision and Pattern Recognition. IEEE, 2016, pp. 770–778.
[3] Ilic, F. et al., "Automated Grain Density Estimation in Austenitic and Martensitic Alloys", 56. Metallographie-Tagung (Materialographie 2022), Saarbrücken, 2022.
[4] Ojala, T. et al., "Performance evaluation of texture measures with classification based on Kullback discrimination of distributions", in Proceedings of the 12th IAPR International Conference on Pattern Recognition, 1994, vol. 1, pp. 582 - 585.
Acknowledgement:
This work was supported by Land Steiermark within the research initiative “Digital Material Valley Styria”.