Iban Plaited Mat Motif Classification using Adaptive Smoothing

• Main Author: Silvia, Joseph
• Format: Thesis
• Language: English
• Published: UNIMAS 2024
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/45964/1/Iban%20Plaited%20Mat%20Motif%20Classification%20Using%20Adaptive%20Smoothing_PhD%20THESIS%20FINALREVIEW%20SilviaJoseph_sign.pdf; http://ir.unimas.my/id/eprint/45964/
• Institution: Universiti Malaysia Sarawak

Iban plaited mats hold a significant cultural importance amongst Borneo’s indigenous communities. These mats are intricately crafted with culturally significant patterns, incorporating a mix of diagonal, symmetrical, geometric, and non-geometric designs. Classifying these motifs poses a challenge due to their complex stylization and incorporation of multiple smaller patterns surrounding the main motif. This work introduces a new image dataset comprising of cleaned, cropped, and perspective corrected Iban plaited mat motif images. First, the accuracy of the Scale Invariant Feature Transform (SIFT) combined with the Random Sample Consensus (RANSAC) algorithm was assessed on the dataset. The optimal peak threshold value for SIFT is reported to be 2.0e-2, and it achieved 100.0% matching accuracy for scale and rotation set. Next, additional local feature descriptors, combined with RANSAC, were evaluated on an expanded dataset. BRISK+RANSAC emerged as the top performer, achieving an average matching accuracy of 85.4%. BRISK+RANSAC encountered challenges in handling unresolved classes in sets with blur, illumination, and viewpoint changes. This is attributed to the spurious classification of smaller motifs affecting the label assignment. To address this issue, an improved classification method with adaptive smoothing is proposed. This method utilizes dynamic thresholds for edge detection and uses morphological operations to enhance the smoothed edges. The proposed method attained 100.0% accuracy for several sets and obtained an average increase of 10.2% accuracy over the baseline method.