Object contour segmentation and representation

• Main Authors: Hu Wu-Chih, 胡武誌
• Other Authors: Hsin-Teng Sheu
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/42225141488529374123

博士 === 國立臺灣科技大學 === 電機工程技術研究所 === 86 === The purpose of this research is to develop 2D object representation and invariants to facilitate data storage, transmission and object recognition. In 2D object representation, the feature extraction is mainly considered to develop in this research, in which the feature extraction is divided into two categories: the local feature and the global feature. The former includes the corner and primitives (such as line segments, circular arcs, elliptic arcs, parabolic arcs, hyperbolic arcs, B-splines and Bezier etc.) which obtain the local feature of curve description. The geometric figures, such as rectangle, circles, ellipses and superellipses etc., that can describe the broad outlines of objects are included in the latter, where the superellipse is a flexible representation which can represent a wide variety of shapes such as rectangles, circles and ellipses etc. In this research, a rotationally invariant two-phase scheme is proposed to detect corners precisely and efficiently. Beside, the corner detection is threshold-free. In the multiprimitive segmentation, the breakpoints obtained using the above scheme are classified as suitable types based on the two-level breakpoint classification which includes an adaptive k-curvature function and a projective height function proposed in this research. Then, a joint tuning procedure is proposed to achieve more accurate segmentation. With the classified breakpoints, curve segmentation is more accurate and reliable. In the extraction of global features, a closed-form solution of the superellipse is derived in this research based on the geometry. The center and the orientation are determined by approximating the superellipse using Fourier description to the first harmonic followed by the consistent symmetric axis method. Then, these parameter are used to find the lengths of the major and the minor axes. Finally, the squareness is solved from the diagonal segment in the superellipse. The proposed scheme is threshold-free and no high curvature bias problem is involved. Besides, incomplete superellipses and pinched superellipses both are extracted satisfactorily. In the development of invariants, the mean amplitude invariants are developedbased on the wavelet descriptors which have the properties of global and localdescription. Besides, thee invariant is not influenced by the transformation of rotation, translation and scaling. The proposed invariants involve the global and local information hence is more reliable for object recognition. The proposed schemes in this research are all efficient and accurate hence the expensive computation in image process can be improved effectively. Besides, the results using the proposed schemes are all threshold-free, thus the human interference can be reduced greatly to elevate the performance of the automatic. Finally, the experimental results reveal the accuracy, the reliability and the practicability of the theorems and the algorithms proposed herein.