T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
the recycled pulp industry [9], [10]. However, traditional methods for fibrillation analysis often performed manually by examining pulp samples under optical microscopy. It is labor-intensive, error-prone, and time-consuming and limit their scalability for real-world industrial applications. This emphasizes the urgent need for automated solutions [11]. Existing commercial analyzers, such as the MorFi Ana- lyzer [12], Metso Pulp Analyzer [13], and Kajaani Fiber Image Analyzer [14], are widely used for measuring general fiber properties like length, width, and coarseness. However, these systems lack the resolution or functionality to directly detect and measure fibrils. Researchers often rely on indirect methods for estimation of fibrillation index, through proxies like surface area, drainage time, or light scattering properties. These approaches introduce calibration challenges and fail to address the need for precise fibrillation index computation at a microscopic level. Recent advancements in image processing present promis- ing opportunities for automating fibril detection and directly measuring the fibrillation index. However, significant chal- lenges persist, including noise in microscopic images, low contrast between fibrils and the background, and the vari- ability in fibril morphology. These limitations hinder existing image processing techniques, restricting their detection capabilities primarily to the fiber level. Addressing these challenges requires a more robust and sophisticated approach capable of accurately identifying fibrils under diverse and complex conditions. Advancements in computer vision and machine learning, particularly convolutional neural networks (CNNs), provide transformative opportunities to overcome these limitations. CNNs have demonstrated remarkable success in detecting intricate patterns and extracting meaningful features from complex visual datasets. However, the detection of very small objects with non-uniform features, such as fibrils, remains a significant challenge [15], [16]. Particularly when small objects coexist with larger, more dominant structures (as in the case of fibrils and fibers). The inherent design of CNNs, which prioritizes the detection of prominent features within an image, often results in weakening signals from smaller structures. As convolutional layers process spatial information through successive stages, the resolution of feature maps diminishes, leading to the potential loss of fine details crucial for small object detection in the deeper layers of the network. Noise and artifacts are also a challenge in small object detection of CNNs. Due to very small objects often visually resemble noise or artifacts, CNNs struggle to differentiate true features from them, leading to a misclassification of noise as objects of interest. This issue also arises in our work. Background noises and textures of fines and fibers pose characteristics visually resemble fibrils and are frequently misclassified as fibrils, resulting in a high false positive rate and overestimation of fibrillation index. These challenges inspire an integrated approach that com- bines the strengths of convolutional neural networks (CNNs)
with traditional image processing techniques. We propose a novel methodology that integrates adaptive image processing with CNNs for automated fibril detection and fibrillation index computation. The approach employs patch-based analysis to handle heterogeneous image conditions and leverages the CNNs’ ability to learn intricate patterns, effectively filtering out irrelevant and noisy patches. This refinement simplifies the subsequent application of adaptive image processing techniques, enabling more accurate and efficient fibril detection. As a result, the method achieves high precision in fibrillation index computation. Experimental results highlight the effectiveness of the proposed method in overcoming key challenges in fibril detection, providing a fast, scalable, and robust solution suitable for industrial applications. The key contributions of this study are as follows: 1) We propose a novel methodology for automated fibril detection and fibrillation index computation in microscopic images. This approach builds on the gold-standard technique of examining pulp samples under optical microscopy for fibril density analysis. This is essential for optimizing both the performance of paper products and the efficiency of the paper manufacturing process. 2) We introduce an innovative approach that combines the strengths of traditional image processing with deep learning techniques to overcome their individual lim- itations. This integrated method effectively addresses key challenges in microscopic fibril analysis and significantly improves detection accuracy compared to results obtained using either technique alone. Through comprehensive evaluation and validation with expert-labeled ground truth data, the proposed method demonstrates strong robustness in real-world applications. 3) Owing to its simplicity and computational efficiency, the proposed method is well-suited for real-time quality control in the pulp and paper industry. The remainder of this paper is structured as follows. Section II reviews existing methods for fibrillation analy- sis, discussing their strengths and limitations. Section III introduces the proposed methodology. Section IV presents the experimental setup, results, and evaluation. Section V concludes the paper by summarizing key findings and outlining potential directions for future research. II. RELATED WORKS Microfibrils and nanofibrils offer significant advantages in papermaking, particularly in strength enhancement and weight reduction, making them valuable for sustainable and high-performance paper products. Their effective application requires addressing optimal fibrillation levels. Therefore, analysis of fibrillation in pulp fibers have been the subject of extensive research, leveraging both image processing and machine learning techniques. However, existing approaches
74635
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease