Comparison of Two Different Proposed Feature Vectors for Classification of Complex Image

Many applications of pattern recognition use a set of local features for recognition .purpose. Instead of using only local features, this paper presents a method to extract features from image body globally as well. The system takes into account several geometrical effects such as area, Euclidean distance etc and their different ratios. It utilizes thresholding and region extraction methods for gray level trademarks images, which furnish these images and segment their separate portions. Thus both local and global traits are constructed that take advantage of the pixel statistics to form a more compact representation of the image, while maintaining good recognition accuracies. Two feature vectors have been proposed. These feature vectors are comprised of nine and seven constituents, respectively. Formation of individual features is very simple involving uncomplicated ratios of geometric and numeric estimate of images pixels. The vectors designed are based on the invariance properlies of individual features. One feature vector is invariant to rotation, translation and size, while the other has an extra invariance regarding scale. In addition, a comparative study on two feature sets is described using backpropagation neural network (BPN) as a classifier.The classification results are encouraging which ranges from 74 to 94%for different data sets.