The study presents a novel hybrid deep learning model for multi-class classification by combining the strengths of two well-established deep-learning architectures: ResNet50 and InceptionV3. The model leverages the complementary features extracted by these two networks by merging their outputs before feeding them into a series of dense layers for final classification. This hybrid approach aims to capture a broader range of representations, overcoming the limitations of single-stream models. Unlike conventional single stream approaches the proposed framework performs parallel feature extraction and feature-level fusion, enabling the model to learn both fine-graded and multiscale discriminative representations more effectively. Experimental evaluation on benchmark datasets confirms that the proposed model outperforms the conventional approaches, additionally the model was extensively compared with recent state-of-the-art deep learning models to validate its effectiveness and superiority. Our hybrid architecture consistently outperformed by achieving higher accuracy and robustness in classifying complex multi-class data and has great potential to minimize misclassification rates and enhance classification robustness.
Sushanki et al. (Sun,) studied this question.