T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
FIGURE 2. (a) an original input image and (b) a fiber segmentation result (a fiber mask image) and (c) an unwanted-object mask image.
and fibrils, low contrast, and variability in fibril morphology. By integrating traditional image processing techniques with deep learning, the methodology achieves a robust and scal- able solution for accurate fibrillation analysis. The process comprises of three main stages: fiber segmentation, fibril detection, and postprocessing, followed by the computation of the fibrillation index. Each step is designed to tackle specific challenges and enhance detection accuracy. A. FIBER SEGMENTATION Fiber segmentation isolates fibers pixels from the background by starting with steps of preprocessing. An input RGB image is first converted to grayscale, and followed by applied adaptive histogram equalization to enhance contrast between fibers and image background. This step ensures that fibers are more distinguishable, thereby improving the reliability of the segmentation process. Fiber segmentation is then performed using Bradley’s adaptive thresholding algorithm, which dynamically adjusts to local variations in illumination, making it robust to uneven illumination. The output is a binary image, where pixels with a value of ‘‘1’’ represent fiber regions, and pixels with a value of ‘‘0’’ denote the background. The initial segmentation result includes fines attached to fibers, individual floating fines, and gaps within the fiber regions (Figure 2(b)). These gaps are caused by the misclassification of the translucent texture of fibers as background during the thresholding process. Postprocessing is required to refine the result to ensure the integrity of the fiber regions, and to provide a clean input for the fibril detection process and the computation of the fibrillation index. To remove fines attached to fibers and individual floating fines, morphological operations and connected-component analysis techniques are employed. First, attached fines are separated from fibers into individual fines using a morpho- logical opening operation with 3-pixel horizontal and vertical line structuring elements. Next, individual fines are removed through connected-component analysis, based on their size and shape. Small connected components with an area smaller than a predefined threshold, area max , or components with a round shape—indicating they are unlikely to be fibers— are excluded from the segmentation result, as shown in Figure 2(c).
Finally, gaps within the fiber regions must be filled to eliminate ambiguity during fibril detection and ensure that no false fibrils are detected in these areas. To remove gaps within the fibers, a morphological fill-hole operator is utilized. To prevent the closure of gaps caused by crossing fibers, which could inadvertently remove fibrils in these areas, the size of gaps eligible for filling is restricted to a predefined maximum threshold, gap max . B. FIBRIL DETECTION Fibril detection is designed to overcome the challenges of low contrast, noise, and fibril-background similarity. This stage employs a patch-based approach and the advanced pattern recognition capabilities of CNNs. The process includes three main steps: image patch categorization, fibril patch selection, and thresholding-based fibril segmentation. STEP1: Image patch categorization The input image is first divided into non-overlapping patches of size20 × 30 pixels to handle impact of uneven illumination and emphasize localized structures of small fibrils. The divided image patches contain different types of data. Patches may consist of only fibers, only fibrils, only noise, fibrils with fibers, fibrils and noise, or plain background. In the task of fibril detection, we will focus exclusively on patches containing fibrils. These include patches with only fibrils, patches with fibrils attached to fibers, and patches containing both fibrils and noise. Since detecting fibrils in different context—whether they are present with fibers or with noise—introduces the distinct challenges. We thus classify the image patches into two categories: patches with fibers and patches without fibers . The classification simply performs by using the fiber segmentation results (Figure 2(b)) as a mask image. Patches overlapping with fiber regions in the mask are categorized as patches with fibers, while the remaining patches are classified as patches without fibers. STEP2: Fibril patch selection Fibril patch selection for each category is carried out using different processes due to the distinct challenges associated with identifying fibrils in each category. A combination of statistical analysis and CNNs are employed, to ensure that only patches containing fibrils are processed further. The diagram in Figure 3 shows the workflow of this step.
74638
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease