転移学習ベースの分類器の適応アンサンブルのための生物学的インスピレーションを受けた象の群れの最適化ベースの方法

Transfer learning has become an important method for image classification when training data is limited. This paper introduces a novel method to build an adaptive ensemble of transfer learning-based classifiers by employing Elephant Herd Optimization (EHO) to enhance image classification performance. Initially 'n' classifiers are built using transfer learning method, then their probabilistic outputs are combined into a single feature matrix. Afterward EHO is used to reveal which classifiers yield maximum contribution to the final decision. These discovered classifiers are then utilized to form ensemble of classifiers. The primary contributions of the proposed methodology include:•Reducing duplication and improving image classification accuracy by utilizing bio-inspired EHO based method to adaptively choose the most efficient subset of transfer learning-based classifiers•Method to build a combined feature matrix by combining probability outputs from several classifiers, which enables the ensemble of classifiers to function on richer, decision-level features.Experiments performed on benchmarked GAIT image dataset and Ocular Disease detection ODIR-5K dataset indicates that this method outperforms classical ensemble strategies, enhancing both accuracy and efficiency.