T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
FIGURE 1. Microscopic image of pulp fibers containing various components. Fibrils ( 20nm - 0.2 μ m [17]) are features that requires ultra-high-resolution images, and advanced specific method for detection. Only fibers (>30 μ m [17]) and fibrils attached to the fibers (highlighted in blue) are included in the fibrillation index calculation.
exhibit notable limitations in accurately and efficiently measuring fibrillation indices, particularly in industrial need. This section reviews these methods, emphasizing their strengths, challenges, and relevance to fibril detection. A. COMMERCIAL FIBER ANALYZERS Fiber analyzer is a necessary tool that enables manufacturers to tailor and optimize a refining process to achieve the desired pulp properties. Commercially available fiber analyzers, such as the MorFi analyser [12], the Fiber Quality Analyzer (FQA) [18], the L&W Fiber Tester [19], the Metso Pulp Analyzer (Metso MAP) [13], and Kajaani Fiber Image Analyzer (presently Valmet) [14], are widely used for measuring fiber properties, e.g. length, width, and coarseness [13], [14], [20], [21]. These modern fiber analyzer systems typically detect crill content by using different wavelengths of light transmitted through a pulp suspension with ultra-high- definition optics. The Valmet Fiber Image Analyzer-Valmet FS5 [14], [22] guarantees to be able to detect particles as small as approximately 1 μ m, which is, however, much larger than microfibrils ( 0.2 μ m). Its resolution limitations mean that some finer fibrillar fines may go undetected. Moreover, no system explicitly documented the capability of direct measurement of fibrillation index from fibril areas. B. RELATED WORKS Microscopy played a pivotal tool in providing direct infor- mation on fiber morphology, especially nanoscale fibril analysis. Image processing methods, such as edge detec- tion and skeletonization, have been extensively applied in fiber morphology analysis. Techniques like the directional wavelet transform [23], [24] and Curvelet transform [25]
have enhanced edge detection and reduced noise in fiber images. However, these methods struggle with complex fibril structures due to challenges like low contrast, noise, and fiber intersections. Skeleton tracing algorithms [26], [27], polynomial modelling [28] and the Radon transform [29] approaches have further improved morphological analysis, but their performance remains limited in detecting fibrils’ intricate patterns, especially in noisy or low-contrast images. Kangas et al. [30] provided an intensive review and outline the effective methods for fiber characterization using different kinds of microscopy, such as SEM, TEM and AFM. Each method requires image processing software to complement the detailed insights provided by microscopy (to extract parameters like fibril width and morphology). In recent years machine learning methods have gained traction for analyzing pulp fibers, addressing some of the challenges faced by traditional techniques and significantly improving accuracy of fiber property measurement. For instance, Yin-Ping et al. [31] introduced a two-tiered repository for handling varying levels of fiber complexity. The proposed system combines a traditional expert system framework with a backpropagation (BP) neural network to improve the recognition of complex fiber characteristics, including intersecting, forked, and curled fibers. However, Setting up the initial expert system requires substantial effort to encode domain knowledge and rules, and the reliance on predefined rules in the fleet repository may lead to errors for borderline or ambiguous cases. Almonti et al. [32] developed two how artificial neural network (ANN) models to predict fiber length based on refining parameters such as pulp flow, refining power, filler content, and wear rate. The experimental results demon-
74636
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease