A Novel Shape Classification Approach Based on Branch Length Similarity Entropy

Abstract

This study presents a novel feature vector for shape clustering based on the Branch Length Similarity (BLS) entropy profile, which is invariant to translation, rotation, and scaling, enhancing its effectiveness for shape analysis. The methodology consists of two steps: the t-distributed Stochastic Neighbor Embedding (t-SNE) technique, which projects feature vectors into a two-dimensional space, and the k-means algorithm, which groups these points into clusters. We applied this approach to three datasets—MPEG-7, Swedish Leaf, and Heptagon—and evaluated its performance. The results revealed significant geometric similarities within clusters, demonstrating the method’s efficacy. Logistic curve modeling was employed to determine the optimal number of clusters, with the Swedish Leaf dataset achieving the highest clustering score (0.885), followed by scores of 0.76 for the MPEG-7 dataset and 0.731 for the Heptagon dataset. The comparative analysis demonstrated that the proposed method outperforms conventional approaches based on Fourier descriptors or Zernike moments, providing a robust and adaptable solution for shape classification across diverse datasets.

Keywords

• Classification
• data structure
• discrete transforms
• entropy
• shape detection