Enhancing CNN Performance Across Diverse Domains: Exploring the Efficacy of Domain Adversarial Neural Network with the Use of CNN for Improved Generalization in Deep Learning
Vs Gaadha, Amrutha S Dev, Vinitha Panicker J, Aswin Sreerag, Anargha Ranjit · 2024
Convolutional neural networks (CNNs) are foundational tools for image classification in deep learning, with models like InceptionNet, ResNet, MobileNet, and DenseNet excelling in feature extraction and validation. However, their performance can suffer in domains with high variability. Domain Adversarial Neural Networks (DANN) address this issue by reducing domain shift through finding domain-invariant representations. DANNs, by training adversaries, enable CNNs to generalize better across different environments, thus enhancing overall performance. This study explores the efficacy of DANN in boosting CNN performance across varied environments. Initially, we detail each base model, focusing on their architecture and feature extraction capabilities. We then introduce DANN as a domain optimization method, showcasing its ability to learn domain-invariant representations. Extensive testing and comparative analysis reveal that integrating DANN with CNN models significantly outperforms using CNNs alone. The findings confirm DANN’s improved generalizability, effectiveness in mitigating domain shift, and potential for advancing deep learning-based image classification tasks.