T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
FIGURE 7. Examples of fibril detection result (cropped from the original images). The detected fibril pixels are overlaid on the input image as red pixels.
F. DISCUSSION The experimental results highlight the effectiveness of the proposed method in fibril detection and fibrillation index computation. The method demonstrates strong capabilities in addressing challenges such as noise, low contrast, and variations in fibril morphology, including small and irregular structures. It performs reliably on fiber sample images with moderate background noise and appropriate illumination dur- ing imaging. However, some limitations remain, presenting opportunities for further refinement and improvement. These include enhancing performance under conditions of noisy backgrounds or abnormal lighting. 1) INCORRECT FIBRIL DETECTION DUE TO FIBER AND FIBRIL CHARACTERISTICS The primary cause of fibril underdetection is the very low contrast between the fibrils and the background, as illustrated in Figure 8. Attempting to address this issue by increasing contrast can inadvertently emphasize artifacts, leading to a higher rate of false fibril detection. The main artifacts contributing to false fibril detection are the highlights and shadows along fiber edges, as shown in Figure 9. These artifacts are not removed during the masking step because doing so risks eliminating short and small fibrils that are crucial to the analysis. Applying shadow and highlight removal algorithms may also inadvertently remove genuine fibril pixels, and fine-tuning these algorithms to address this issue would significantly increase processing time, making them unsuitable for real-time applications. Another significant factor contributing to high false fibril detection rate (Table 3) is the presence of voids within fibers or fines, as depicted in Figure 10. Some fibers and fines exhibit a relatively transparent texture, which prevents these regions from being classified as fibers during the fiber segmentation process. Consequently, these regions remain unmasked and are included in the fibril detection step. Due to the texture of these regions closely resembles areas where fibrils attach to fibers, portions of the fiber and fine textures are misclassified as fibrils. Addressing this issue within the fiber segmentation process is challenging due to the varying sizes and patterns of voids within fibers and fines. Fine-tuning
FIGURE 8. Underdetected fibrils caused by low contrast between the fibrils and the background.
FIGURE 9. False fibrils caused by highlights along fiber edges.
the segmentation process to account for these complexities would require additional processing time. To address this problem, this work leverages a CNN trained to detect patterns of voids within fibers or fines. The CNN is used to identify
74643
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease