A Hierarchical Multi-Class Classification System for Face and Text Datasets

Ashish Saini¹Nasib Singh Gill¹Preeti Gulia¹KHUSHWANT SINGH¹Fernando Moreira^2*

• ¹Maharshi Dayanand University, Rohtak, Haryana, India
• ²Portucalense University, Porto, Portugal

In the era of rapidly growing multimedia data, the need for robust and efficient classification systems has become critical i.e. identification of class name and pose or style. The paper first provides an understanding of the organization of data, next feature selection (i.e. edge) using the k-means segmentation technique is explained. Further for optimization of features, the linear regression technique is employed. The optimized features can be directly used with classifiers but to reduce the noise outliers are identified and removed from training data. The classifiers involved in training and recognizing the face or text class label. After the prediction of class labels, the distance matrix-based technique is used to identify the style or pose name. Finally, the experiments are conducted with the help of the ORL dataset (40 class and 10 poses in each class) and character dataset (36 characters and 10 font styles in each character). The experimental consequences demonstrate the proposed methodology accurately classifies the hierarchically organized data and prove superiority over KNN and Bayesian based classification in place of SVM. The system provides up to 100% accurate classification outcomes for outlier removed data, and for basic features, it is up to 98%. Unlike traditional flat classification approaches, our system leverages hierarchical structures to enhance classification accuracy, scalability and interpretability.