Developing a Computer Vision Model to Classify Abaca Fibers Using YOLOv8

Abstract

Introduction:  Abaca (Musa textilis), a banana species endemic to the Philippines, which is vital for the country’s economy, supplying 87% of the global market. T he industry faces significant challenges in fiber classification, relying on a limited number of trained professionals to visually inspect fibers within a restricted timeframe, leading to inefficiencies and high costs. Although some automated systems using convolutional neural networks (CNN) have been developed, they are limited to processing single fibers at a time.

Objectives: This study aims to improve the classification process of abaca fibers. This study aims to create a deep learning-based classification model utilizing YOLOv8 architecture to accurately distinguish between various grades or types of abaca fibers based on their visual characteristics and conduct laboratory tests to evaluate the effectiveness of the classification and sorting system.

Methods: This study focuses on developing a YOLOv8 model for classifying grades S2 and S3 abaca fibers. Images were collected from two trading centers in Iligan City. These images were manually annotated, with the dataset split into training and validation sets. Then the images undergo preprocessing and augmentation including flipping, cropping, blur, and noise adjustment, which results in a 2,332 training images and 295 test images. The model was then tested on 40 new, unannotated images in a laboratory setting to evaluate its performance in classifying abaca fibers.

Results: The model performed well in detecting and classifying abaca fibers, achieving high precision, recall, mAP50 and mAP50-95. The training process showed positive results, with decreasing box loss in later epochs and a quickly dropping validation loss, indicating effective generalization. Although there were minor signs of overfitting and the losses stabilized around 1.0, the overall convergence of training and validation losses suggested reliable performance. The confusion matrix showed strong classification accuracy, with 136 correct predictions for Grade S2 and 132 for Grade S3 fibers, despite some false positives in images without fibers. In the final testing phase, the model successfully identified 35 out of 36 abaca fibers, demonstrating its strong classification abilities.

Conclusions: In conclusion, the researcher was able to create a model to classify grades S2 and S3 abaca fibers. The trained YOLOv8 model proved to be a robust model in classifying abaca fibers. Since the dataset uses a variety of settings, including the time of day and light source, the model was able to generalize well in classifying abaca fibers.