Invariant Tchebichef And Krawtchouk Moments For Shape Recognition: A Performance Evaluation Study

Abstract

In the last few decades, orthogonal invariant moments have become a considerable technique used in feature extraction. Thanks to their invariability against geometric transformations such as translation, scale changing, and rotation. This property makes them very useful in shape recognition and image classification tasks. In this paper, an application case and performance evaluation of the invariant moments derived from the known orthogonal Tchebichef and Krawtcouk polynomials are presented. The invariants are computed by creating a relationship between these moments and invariant geometric ones. These invariants are used to form an efficient vector of features. However, for performance evaluation, we adopt two classification models as input. The invariant descriptor vector is considered an input feature vector for classification models. The first model is a k-nearest neighbor, and the second is a neural network model. These models are used for handwritten digit classification in free-noise and different environmental noise types. In addition, the classification with the invariant method is also compared with the convolutional neural network image classifier. The results show the power and efficiency of invariant orthogonal moments in shape and pattern recognition.