Iban plaited mat motif classification with adaptive smoothing

Abstract

<span lang="EN-US">Decorative mats plaited by the Iban communities in Borneo contains motifs that reflect their traditional beliefs. Each motif has its own special meaning and taboos. A typical mat motif contains multiple smaller patterns that surround the main motif hence is likely to cause misclassification. We introduce a classification framework with adaptive sampling to remove smaller features whilst retaining larger (and discriminative) image structures. Canny filter and probabilistic hough transform are gradually applied to a clean greyscale image until a threshold value pertaining to the image’s structural information is reached. Morphological dilation is then applied to improve the appearance of the retained edges. The resulting image is described using binary robust invariant scalable keypoints (BRISK) features with random sample consensus (RANSAC). We reported the classification accuracy against six common image deformations at incremental degrees: scale+rotation, viewpoint, image blur, joint photographic experts group (JPEG) compression, scale and illumination. From our sensitivity analysis, we found the optimal threshold for adaptive smoothing to be 75.0%. The optimal scheme obtained 100.0% accuracy for JPEG compression, illumination, and viewpoint set. Using adaptive smoothing, we achieved an average increase in accuracy of 11.0% compared to the baseline.</span>

Refining disparity maps using deep learning and edge-aware smoothing filter

“…Mismatches are interpolated with median value of nearest matching neighbors, while occlusions use disparity value of nearest matching neighbor on right (or left) or raw disparity value if no match found. The quality of the disparity map is enhanced further through the application of EASF which is a modification of the traditional edge-preserving filter [38], [39] that iteratively applies the filter, smoothing the image while preserving its edges [40]. The idea is that the amount of smoothing between two pixels should depend on how far apart they are.…”