T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
FIGURE 11. Examples of false positives, where patches without fibrils are misclassified as containing fibrils.
FIGURE 10. Texture of fibers that caused false fibrils.
and exclude these regions from the fibril detection results, significantly reducing false fibrils in the output without compromising processing efficiency. 2) INCORRECT FIBRIL DETECTION DUE TO CLASSIFICATION ERRORS In this study, we leverage the capabilities of CNNs to classify complex and diverse patterns, enabling the filtering of patches containing fibrils for accurate fibril detection. Two CNN classifiers are trained to handle patches from two distinct categories. For the first category (Patches Containing Fibers), a ResNet-50 model is employed to differentiate between patches with fibrils attached to fibers (class 1) and patches containing only fibers without fibrils (class 0). Classification errors in this category result in the following cases: 1) False positives, where patches without fibrils are misclas- sified as containing fibrils, leading to false fibril detection (Figure 11). False fibrils arise in this case because the fibril detector attempts to identify fibril pixels where none exist, leading it to mistakenly classify highlighted pixels along the edges of fibers as fibril pixels. 2) False negatives, where patches with fibrils attached to fibers are excluded from fibril detection, contributing to an underdetection rate (Figure 12). For the second category (Patches Without Fibers), a ResNet-50 model is used to classify patches with noise or fine debris (class 0) versus patches with fibrils (class 1). Errors in this category produce similar challenges: 1) False positives, where noise or fine debris are misclassified as fibrils, leading to increasing false fibril rate (Figure 13). 2) False negatives, where patches with fibrils are excluded from fibril detection, resulting in an underdetection rate (Figure 14). By employing ResNet-50 models tailored to these two categories, the method aims to improve the accuracy of fibril detection by reducing noisy fines and artifacts. However, addressing the false positive and false negative rates remains crucial for further enhancing the performance of fibril detection. Training the CNN on a more diverse dataset with challenging background conditions could improve its robustness.
FIGURE 12. Examples of false negatives, where patches with fibrils attached to fibers are misclassified as patches containing only fibers.
FIGURE 13. Examples of false positives, where noise or fine debris patches are misclassified as fibril patches.
FIGURE 14. Examples of false negatives, where patches with fibrils are misclassified as noise or fine debris patches.
3) FIBRIL MORPHOLOGY ANALYSIS While the methodology has some limitations, it marks a significant advancement in automated fibril analysis, offering a robust framework for industrial applications and material science research. Although this paper primarily focuses on fibrillation index calculation, the detected fibrils can also serve as input for further morphology analysis. Using our previously developed skeleton reconstruction algorithm [36], additional automated inspection and measurement tasks can be performed. This algorithm excels at isolating individual objects within clusters of overlapping objects and deal- ing with various crossing patterns, making it particularly well-suited for analyzing the complex characteristics of fibrils in this study. The length and morphology of the fibrils can be measured directly from the reconstructed skeleton.
74644
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease