T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
FIGURE 5. An example of removing false fibrils within a fine using a CNN model. (a) Unfilled gap within a fine (b) False fibrils inside the gap (c) Results after eliminating false fibrils.
Under-detection Rate = Missed Fibril Pixels Total Fibril Pixels in Ground Truth × 100 (4) • Over-detection rate measures the percentage of pixels incorrectly identified as fibrils: Over-detection Rate = False Fibril Pixels Total Fibril Pixels in Ground Truth × 100 (5) 2) FIBRILLATION INDEX ACCURACY The fibrillation index obtained using the proposed method was compared to manual ground truth values to evaluate accuracy. The average magnitude of errors between the manual ground truth and machine-generated results is expressed as the absolute deviation error. This is calculated using the mean absolute deviation (MAD), which represents the average of the absolute values of the errors. MAD = n i = 1 | Manual Value i − Machine Value i | n (6) where n is the total number of data points in the dataset. MAD is an effective metric for assessing the typical size of machine errors, as it focuses solely on the magnitude of the errors, disregarding their direction or polarity. D. EXPERIMENTAL SETTINGS The dataset was divided into training and test sets. The training set comprised 38 images, primarily from the low- dried-200 setting, as this data was available earlier. The test set included 20 images from all refinement conditions to ensure unbiased evaluation. Table 2 summarizes the dataset details. The statistical threshold values used in the proposed method, namely T SD , area max , and gap max , are determined experimentally by performing a statistical analysis on the training images, and are set to 1.2, 500, and, 250, respectively.
number of fiber pixels [8], as shown below : Fibrillation index = Number of fibril pixels Number of fiber pixels ×
100 (3)
IV. EXPERIMENTS AND DISCUSSION A. EXPERIMENTAL SETUP
The experiments were conducted on a system equipped with an AMD Ryzen 7 7735HS processor, NVIDIA RTX 3050 GPU (4GB VRAM), and 16GB RAM. Image processing and model implementation were performed using MATLAB 2024b. B. DATASET PREPARATION The dataset used for this study was sourced from the Department of Forest Products, Faculty of Forestry, Kasetsart University. It consists of microscopic images of pulp fibers, each containing at least three fiber strands, captured under a light microscope at 10 × magnification. The images represent pulp samples refined under varying conditions of intensity (high and low), pulp type (dried and undried), and specific energy consumption (SEC). Two types of refiner plates were used, with samples collected at five SEC levels for each plate. Example of images used in the experiments are shown in Figure 6. Images with significant contamination or staining were excluded from the analysis, ensuring data quality.
C. EVALUATION METHOD 1) FIBRIL DETECTION ACCURACY
Ground truth fibril maps were manually annotated by experts in pulp refining. Each pixel in the ground truth was labeled as either ‘‘fibril’’ or ‘‘background.’’ The proposed method’s performance was evaluated by comparing its fibril detection results against these ground truth images, calculating under-detection and over-detection rates. • Under-detection rate measures the percentage of true fibril pixels not identified by the method:
74641
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease