A customized template matching classification system

Abstract

This paper presents two novel classification techniques, the customized template matching classifier (CTMC) and the dynamic template matching classifier (DTMC), which aim to significantly enhance the performance of the minimum distance classifier (MDC). CTMC tailors a feature subspace specifically for MDC, leveraging a set of effective templates that capture the distinguishing characteristics of each class. This customized feature space ensures accurate representation and enhanced distinguishability between classes, thereby improving MDC’s classification accuracy. DTMC, on the other hand, builds upon the CTMC approach by introducing a dynamic template optimization process. Inspired by semi-supervised learning techniques, DTMC utilizes unlabeled data to enrich class information and iteratively update the templates in the feature space. This dynamic optimization process allows DTMC to adapt to variations in the data, further enhancing the classification performance of MDC. Our key contributions include: (1) introducing the concept of a customized feature space tailored for MDC, demonstrating its effectiveness in improving classifier performance; (2) presenting CTMC and DTMC as comprehensive classification systems that seamlessly integrate feature extraction and classification, outperforming traditional loosely coupled approaches; and (3) incorporating a reliability mechanism to assess the classification of test samples, enabling the selection of high-reliability samples to update class templates, effectively addressing the issue of limited labeled data and further boosting the overall performance of the classification system.