الفهرس 
Abstract
IntroductionOnly 14 pages are availabe for public view
 |  | from | 126 |  |  |
| | | from | 126 | | |
Abstract
The aim of feature extraction in image processing is simplifying the amount of resources required to describe a large set of data accurately by transforming the input data into set of features. To select a set of appropriate numerical features from the interested objects for the purpose of classification has been among the fundamental problems in the pattern recognition system. One of the solutions, the utilization of moments for object characterization has received considerable attentions in recent years. Many applications in image processing are based on feature extraction. In this thesis, we present four contributions to improve image applications that are based on feature extraction.
Firstly, new techniques are described to increase the efficiency of Pseudo Zernike moment (PZM) computation for small images. It is proposed to compute PZM for a decimated image, and use Bspline interpolation to interpolate between image pixels. This significantly improves PZM computations due it is smooth and robust performance. Further improvements are also possible, by basing our computations on least square Bspline decimated images, as well as optimizing some of PZM coefficients.
Secondly, the rotation invariant property of Pseudo Zernike moments is used in modifying a natural preserving watermarking technique, that are known for its robustness against cropping, compression and noise attacks, to make it withstand rotation attacks.
Thirdly, feature extraction is used to improve the fingerprint segmentation, one of the main steps of Automatic Fingerprint Identification System. An accurate segmentation algorithm is helpful in speeding up the computations. In this thesis, a novel segmentation method is presented that are mainly based on block range calculation together with some morphological opening and closing operations to extract the fingerprint foreground. The proposed method provides accurate high-resolution segmentation results compared to other segmentation methods.
Finally, a new approach for fingerprint image enhancement is introduced to improve the extraction of features from poor quality fingerprints. The enhancement method is based on features extracted from the fingerprint such as the local ridge orientation and local frequency. We use two stage for enhancement in both the spatial domain and the frequency domain. The fingerprint image is enhanced in the spatial domain with a directional mean filter or an interpolation technique. Then, a frequency bandpass filter that is separable in the radial and angular frequency domains, is employed. The center frequency of the radial filter is accurately determined using a novel Radon-based technique. The second stage is repeated two or three times. The iteration method enhances gradually and significantly the low quality fingerprint images. We also measure the effectiveness of the enhancement algorithm and show that it can improve the sensitivity and recognition accuracy of existing feature extraction and matching algorithms.