GAN-Augmented Ensemble Models for Wildlife Image Recognition on Caltech‑256 Dataset

Abstract

The correct classification of wildlife images is still a major problem in the light of the low supply of labeled data and the abundance of intra-class variance. This paper will present a powerful ensemble model that includes synthetic data augmentation with Generative Adversarial Networks (GANs) and transfer learning to improve the performance of wildlife recognition. The pipeline is built on a hand-picked sample of the Caltech-256 data, where image preprocessing is performed and Synthetic samples are produced using GAN to add variety to training. Three convolutional neural networks, ResNet50, VGG16 and inceptionV3 are trained on these augmented data and their representational strengths are used to their advantage by utilizing them together. Then a weighted voting committee is created on the basis of the individual model accuracies to create the final prediction output. Experimental results demonstrate that the proposed GAN-augmented ensemble significantly outperforms both traditional augmentation baselines and single-model configurations, achieving an accuracy of 93.29%. The approach highlights the effectiveness of combining generative modeling and ensemble strategies for improved performance in small-sample, high-variability wildlife classification scenarios.