Optimization of Convolutional Neural Networks Through Soft Voting Ensemble and Fine-Tuned Approach to Image Classification

Abstract

This study introduces an optimization method to enhance image classification accuracy using a Soft Voting Ensemble of fine-tuned convolutional neural networks (CNNs). The goal is to evaluate whether combining multiple CNN architectures can outperform individual models. Three pre-trained networks DenseNet, ResNet, and EfficientNet were fine-tuned through transfer learning on a Kaggle image dataset. To improve generalization, data augmentation techniques such as random rotation, flipping, and zooming were applied. After fine-tuning, the models were integrated using a soft voting strategy that averages prediction probabilities to determine the final class. Performance was tested on 94 unseen images, fully separated from the training data. The individual accuracies were 91.5% for DenseNet, 87.2% for ResNet, and 93.6% for EfficientNet. The proposed ensemble achieved the highest accuracy of 94.7%, with precision, recall, and F1-score all reaching 0.95. These findings indicate that the ensemble approach successfully combines the strengths of different CNNs, reduces classification errors, and increases model robustness. Overall, the soft voting ensemble provides a reliable, scalable, and effective solution for improving CNN-based image classification, especially when dealing with limited or diverse datasets