Image feature extraction invariant to color intensity, rotation and scaling using Eigenvector-guided self-organizing mapping neural network

Abstract

This paper proposes a new feature extraction algorithm to extract the invariant features of an image based on the concept of Principle Component Analysis (PCA) and competitive learning algorithm. PCA assists in guiding the rotation axis of the image but it cannot find the exact rotational direction of the image if the image is rotated through an angle greater than 180 degrees. Self-partitioning competitive learning can capture the texture of an image by using the location of weight vectors and it embeds the invariant ability of scaling. The proposed algorithm is developed to apply to colored-texture images of size at least 256x256 pixels. In addition to scaling and rotation invariance, color intensity change are addressed in the paper. The experiment data contains 34 color-textured images including images having the same shape but different colored texture and having the similar color distribution but different shape. Experimental results show that our new algorithm performs the discriminating capability of images having the same shape but different colored texture and also reserves the invariant ability.