Advancing Convolutional Neural Networks: Novel Techniques and Evaluations for Enhanced Robustness and Generalization

Abstract

This thesis advances the field of computer vision by addressing the critical challenges of out-of-distribution (OOD) generalization and robustness. Computer vision systems, particularly those based on convolutional neural networks (CNNs), have shown remarkable performance in controlled settings with in-distribution data. However, their ability to generalize to new, unseen data that differs significantly from the training set remains a substantial obstacle. This issue is especially pertinent in real-world applications where models frequently encounter data with varying conditions, such as changes in lighting, occlusions, or entirely new objects and scenarios. Ensuring that these systems can maintain high performance and reliability under such diverse conditions is crucial for their deployment in critical areas like autonomous driving, healthcare, and surveillance. Through a comprehensive study, this research explores innovative methodologies to enhance the robustness and generalization capabilities of CNNs, providing practical solutions to these challenges. To address this issue, we introduce four novel contributions to the field. First, the "noisy/noise-free" (NNF) training method improves the generalization capabilities of CNNs by incorporating a balanced mix of noisy and noise-free images in the training set, significantly enhancing overall generalization and classification accuracy on noisy datasets. Second, we present Dominant Feature Masking (DFM), a data augmentation approach that strategically conceals dominant features within images to encourage the network to learn both primary and secondary attributes, thus improving OOD prediction performance. Third, we develop the Versatile Evaluation Benchmark (VEB), a comprehensive evaluation framework that rigorously tests the domain generalization and OOD performance of CNNs across various challenging scenarios. Fourth, the thesis introduces the Dynamic Attention Layer (DAL), a novel layer that dynamically adjusts attention weights based on feature relevance during training, allowing CNNs to focus on both dominant and secondary features, thereby enhancing their robustness and adaptability. Our findings demonstrate significant improvements in the generalization performance of CNNs, particularly in handling diverse and unpredictable scenarios. The research offers valuable insights and practical solutions for developing more robust and reliable computer vision systems.