<![CDATA[This study aims to evaluate the effectiveness of multiple feature representations in classifying Pikachu images into three distinct visual categories: anime, action figures, and hand-drawn illustrations. The primary challenge lies in limited data availability and the high visual variability across styles, resulting in significant inter-class similarity and intra-class diversity. To address this issue, the study employs a transfer learning approach utilizing pre-trained Convolutional Neural Networks (CNNs), namely VGG-16 and Inception-V3, alongside painterly feature descriptors. The dataset comprises 351 images collected from open-access sources with balanced class distribution. Extracted features are subsequently classified using Support Vector Machines (SVM) and shallow Neural Networks. The findings demonstrate that integrating deep semantic features with artistic representations significantly improves classification accuracy compared to single-feature approaches. These results highlight the critical role of hybrid feature engineering and classifier selection in achieving robust image classification performance under data-constrained conditions.]]>
Copyrights © 2026
