T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
TABLE 1. Summarization of the existing methods reviewed in the paper.
strated that by Employed statistical techniques like principal component analysis (PCA) and hierarchical clustering (HC) to reduce noise and identify outliers can improve accuracy and reliability of the model. The proposed method allows for predicting fiber length in real time, however, the models were tailored to specific refiners and operational parameters, limiting their direct applicability to other setups without retraining. Kontschieder et al. [33] proposed the method for detecting paper fiber cross sections in microtomy images. The method leverages a novel discriminative edge fragment descriptor that captures angular relations between fragment points, and uses a Random Forest classifier trained on these descriptors. The system predicts fiber cross-section locations through a generalized Hough voting scheme, allowing robust detection of fiber cross sections, including those with significant shape deformations, collapsed structures, and variable sizes. The proposed method successfully detected fibers with significant shape variations and collapsed structures, however, it relies heavily on accurate edge detection. Poor edge extraction (due to noise or low contrast) can degrade detection performance.
Lindström et al. [34] explored four machine learning models - Lasso Regression, Support Vector Machine (SVM), Feed-Forward Neural Networks (FFNN), Recurrent Neural Networks (RNN) to enhance the classification of pulp parti- cles. By Incorporating In-House Image Analysis Software for Extracting additional parameters (e.g., light attenuation) and Preprocessing transformations like Yeo-Johnson and binary transformations, Lasso Regression achieved the highest classification accuracy, RNNs exhibited Instability with significant variability in performance. While significant progress has been made in fiber analysis, no existing methods provide a comprehensive, automated solution for fibril detection in pulp fibers. This study bridges this gap by integrating adaptive image processing with CNNs to deliver accurate, scalable, and efficient fibrillation index computation. Table 1 summarizes the existing methods reviewed in this paper. III. THE PROPOSED METHOD This study introduces a novel methodology for automated fibril detection and fibrillation index computation, addressing the challenges such as noise, size disparity between fibers
74637
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease