T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
FIGURE 6. Example of images used in the experiments.
TABLE 3. Comparison of fibril quantities between program and manual methods.
TABLE 2. Details of the dataset used in the experiments.
These threshold are recommended to determine based on data and applications that implement this method. The ResNet-50 models used in this study are retrained with a stochastic gradient descent, momentum as the optimization function, a maximum of 6 epochs, and an initial learning rate of 0.003. Four-fold cross-validation is employed to validate the accuracy of the model. The YOLOv4 models used in this study were retrained using the Adam optimizer, with a maximum of 70 epochs and an initial learning rate of 0.001. L2 regularization was applied with a value of 0.0005, and the data was shuffled at each epoch to enhance model performance.
detected by the algorithm. The results highlight the method’s capability to identify fibrils with diverse characteristics, ranging from very short fibrils (2)-3 pixels in length) to long, straight, or curvilinear threads. By combining patch- based analysis with deep learning, the method reduces the impact of global noise and focuses on localized regions, enabling accurate detection even in challenging conditions, such as areas with heterogeneous structures, irregular fiber textures, and low contrast. Additionally, preprocessing and postprocessing steps play a critical role in mitigating issues like background blending and contaminant interference, further enhancing the overall detection performance. 2) FIBRILLATION INDEX RESULTS The fibrillation indices calculated by the proposed method demonstrate strong correlation between manual ground truth and machine measurements. Table 3 compares fibrillation index values across 20 test images, with an average absolute deviation of 0.449%, underscores the method’s reliability for real-world applications.
E. EXPERIMENTAL RESULTS 1) FIBRIL DETECTION RESULTS
The proposed method demonstrated promising performance of fibril detection across diverse refining conditions. Table 3 summarizes the under-detection and over-detection rates for the 20 test images. On average, the under-detection rate was 18.76%( ± 12.21), while the over-detection rate was 14.32% ( ± 10.57). Images with low contrast and heavy noise showed relatively higher error rates (Table 3), as some fibrils blended into the background or were obscured by contaminants. Examples of successful fibril detection are illustrated in Figure 7, where several challenges leading to misdetection were mitigated through preprocessing and postprocessing steps. The image is cropped from the original image and enlarged for clarity, to make the fibrils more visible. Red pixels overlaid on the input image indicate the fibrils
74642
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease