Browsing by Author "Aukkapinyo K."

Filter results by typing the first few letters
Now showing 1 - 5 of 5
  • No Thumbnail Available
    ItemMetadata only
    Automated tongue segmentation using deep encoder-decoder model
    (2023-01-01) Kusakunniran W.; Borwarnginn P.; Imaromkul T.; Aukkapinyo K.; Thongkanchorn K.; Wattanadhirach D.; Mongkolluksamee S.; Thammasudjarit R.; Ritthipravat P.; Tuakta P.; Benjapornlert P.; Mahidol University
    This paper proposes a solution of tongue segmentation in images. The solution relies on a convolutional neural network, using deep U-Net with deep layers of encoder-decoder modules. The model is trained with a starting resolution of 512 x 512 pixels. To enhance the segmentation performances of the trained model across recording environments, three main types of data augmentations are added in the training process, including additive gaussian noise, multiply and add to brightness, and change color temperature. They could also handle an inadequate number of data samples in the limited datasets. The proposed method is evaluated based on four measurement metrics of Dice coefficient, mean IoU, Jaccard distance, and accuracy. The model is successfully trained on publicly available datasets, and then transferred to be tested with the self-collected dataset in the real-world environment.
  • No Thumbnail Available
    ItemMetadata only
    Automatic classification of mangosteens and ripe status in images using deep learning based approaches
    (2023-01-01) Kusakunniran W.; Imaromkul T.; Aukkapinyo K.; Thongkanchorn K.; Somsong P.; Mahidol University
    After the mangosteen-harvest, it is necessary to process a grading assessment which contains a key step of classifying a ripe status. This paper aims to develop an automatic solution to replace a manual process by human experts for classifying mangosteen and their ripe status in images. The classification solutions are developed based on deep learning techniques. These classification models are constructed by attempting on four architectures (i.e. DenseNet, EfficientNet, ResNet, and VGG) of convolutional neural networks (CNN). The models are trained using well-known and new prepared datasets. Two training strategies of multi-class and binary classifications are attempted in our experiments for distinguishing mangosteen from other fruits. It is reported that the multi-class classification performs better than the binary classification, with the precision, recall, and f1-score of 100%. In addition, a gradient-weighted class activation mapping (Grad-CAM) is used to demonstrate the reliability of the trained models. The proposed solution based on EfficientNetB0 performs the best for classification of mangosteens and their ripe statuses with the average accuracies of 100% and 98% respectively. The multi-class CNN-based classification is developed for solving a real-world problem of the ripe status classification. Alternative CNN architectures are attempted for finding the best-fit solution on a publicly available dataset and a self-collected dataset from a web scraping. The computed heatmaps show that it is not necessary to perform the mangosteen segmentation, the classification task could be performed directly where background and irrelevant parts of images are not/or less used.
  • No Thumbnail Available
    ItemMetadata only
    Detection of translucent flesh disorder and automatic grading of mangosteens in multi-view images
    (2025-01-01) Kusakunniran W.; Imaromkul T.; Aukkapinyo K.; Thongkanchorn K.; Somsong P.; Tiyayon P.; Kusakunniran W.; Mahidol University
    In this paper, convolutional neural network (CNN)-based solutions are developed for grading assessment and flesh disorder detection of mangosteens in images. The grading is set to three classes of three quality levels based on the local market, where the data were collected. In addition, three flesh disorders/status are focused in this work, including translucent flesh disorder, gamboge, and rotten. Three types of solutions are attempted in this paper. The first solution relies on the well-known CNN architectures with the transfer learning and data augmentation. The second solution is developed based on the detection model, i.e., YOLOv8. The third solution is to design a new architecture by taking into account of human expert knowledge that is used for the manual grading and detection. Multiple views of each mangosteen must be considered simultaneously for the disorder detection. Four side views should be considered together, before looking at the top and bottom views. This is a very difficult task even for the human experts. The proposed solutions are trained and evaluated on the self-collected dataset of 206 mangosteens captured under six views (i.e., top view, bottom view, and four side views). The proposed solutions could achieve the perfect accuracy of 100% for the grading and up to 78% AUC for the disorder detection.
  • No Thumbnail Available
    ItemMetadata only
    Manga Face Detection on Various Drawing Styles Using Region Proposals-Based CNN
    (2023-01-01) Aukkapinyo K.; Mahidol University
    Faces of characters in comic books can be used as meta-features for manga analytics. Manga character faces are not easy for a machine to detect when compared to human faces due to the high variation of drawing styles from various distinct authors. There exist several convolutional neural network-based (CNN-based) frameworks that can achieve high accu-racy in an object detection task. However, their drawback is time and resource consuming to perform data modeling due to the nature of deep learning. Thus, this paper is to propose a method to develop a model using Mask R-CNN, which is one of the CNN-based frameworks, with the transfer learning technique in order to reduce training time and resources while main-taining high performance in the manga character face detection task. The proposed method could achieve the average precision of 87% in the manga character face detection tasks on both seen and unseen drawing styles. It significantly outperforms the existing conventional methods. Moreover, pre-trained weights from MS COCO dataset are transferable to manga character face detection tasks. Therefore, a well-performed manga character face detector could be developed using a limited amount of training data and time.
  • No Thumbnail Available
    ItemMetadata only
    Measurement of Tongue Motion using Optical Flows on Segmented Areas
    (2022-01-01) Kusakunniran W.; Aukkapinyo K.; Borwarnginnn P.; Imaromkul T.; Thongkanchorn K.; Wattanadhirach D.; Mongkolluksamee S.; Thammasudjarit R.; Ritthipravat P.; Tuakta P.; Benjapornlert P.; Mahidol University
    A trajectory of the tongue has several benefits in various domains such as articulatory and medical. It allows a user to analyze human speech or diagnose anomaly tongue movement of patients. This research focuses on estimating tongue motion. Most existing solutions apply traditional image processing techniques to a sequence of images to compute motion. Although they can precisely estimate a tongue motion, there are drawbacks to practicality and scalability. It is because of the high cost of medical imaging devices such as magnetic resonance imaging (MRI) and ultrasound scanners. There is also overhead in the preparation of marking on the face of the patient. On the other hand, the optical How algorithm can produce motion vectors on videos obtained from a commercial camera. This paper proposes a solution that can estimate tongue motion with more praetieality and less overhead. An average motion vector can be precisely computed within a region of interest of a tongne.