| Most recognition systems can be divided into four stages: data acquisition, feature extraction, matching and verification. Given that the objects of interest occupy only a small portion of the image, most of the data present in the image is extraneous. This makes feature extraction a very difficult time consuming and computationally intensive task. The task is further complicated because it is not easy to define what consists of a feature. The definition of a feature is directly related to the way each of the objects are represented and processed by the matching algorithms. Traditionally, linear and curvilinear segments, conics, etc. have been used as features. Object representation can be generated by combining various features together. Geometric information is derived from the extracted features. Efficient representation of the extracted features and their properties plays a very important role in designing a feasible, fast and reliable matching based object recognition system. This dissertation outlines algorithms based on Hough transform and its variations for extraction of 144 linear features from 2-dimensional grayscale images. Techniques to determine length and spatial positioning of the extracted linear features are studied. The concept of parallelism has been introduced to the feature extraction process. The research also proposes a massively parallel object representation scheme which can be exploited for fast recognition of objects. |
