ADVANCES IN IMAGE-BASED DATA HIDING, FEATURE DETECTION, GRID ALIGNMENT, AND DOCUMENT CLASSIFICATION
Data embedding tools such as barcodes are very popular nowadays but not aesthetically pleasing. In this research, we propose a watermarking scheme and an image-based surface coding scheme using the grid points as fiducial markers and the shifted points as data-bearing features. Detecting and aligning point grids play a fundamental role in these applications. Joint determination of non-grid points and estimation of non-linear spatial distortions applied to the grid is a key challenge for grid alignment. We modify a SIFT-based surface feature detection method to eliminate as many spurious feature points as possible and propose a grid alignment algorithm that starts from a small nearly regular region found in the point set and then expands the list of candidate points included in the grid. Our method is tested on both synthetically generated and real samples. Furthermore, we extend some applications of the surface coding scheme to 3D space, including hyper-conformal mapping of the grid pattern onto the 3D models, 3D surface feature detection, and 3D grid points alignment.
A document routing system is crucial to the concept of the smart office. We abstract it as an online class-incremental image classification problem. There are two kinds of classifiers to solve this problem: exemplar, and parametric classifiers. The architecture of exemplar-based classification is summarized here. We propose a one-versus-rest parametric classifier and four different updating algorithms based on the passive-aggressiveness algorithm. An adaptive thresholding method is also proposed to indicate the low-confidence prediction. We test our methods on 547 real document images that we collected and labeled and high cumulative accuracy is reported.
History
Degree Type
- Doctor of Philosophy
Department
- Electrical and Computer Engineering
Campus location
- West Lafayette