Signal to noise dynamics and feature robustness in convolutional neural networks

Abstract

Convolutional neural networks (CNNs) have become foundational to modern computer vision owing to their ability to learn hierarchical feature representations from raw data. Despite their widespread success, the internal relationship between feature quality and the underlying signal to noise ratio (SNR) within these networks remains insufficiently understood. This gap is significant, as imagery data is frequently affected by noise, acquisition artefacts, and heterogeneous backgrounds that challenge the reliability of deep learning models. This thesis presents a systematic investigation into how SNR evolves across the layers of several widely used CNN architectures and examines how these dynamics relate to discriminative performance. Using a dataset of brain tumour Magnetic Resonance Imaging (MRI) slices with tumour masks, the study isolates signal and background components to quantify layer-wise SNR while evaluating representational quality through lightweight linear probe classifiers. Two scenarios are analysed: natural background noise inherent to the dataset and controlled Gaussian noise injected at varying intensities. The results reveal that robustness to noisy inputs is strongly dependent on architectural design. Feed-forward models such as Visual Geometry Group 16-layer network (VGG16) and Visual Geometry Group 19-layer network (VGG19) show a direct correspondence between SNR and classification accuracy, with deeper layers recovering informative features even when early representations degrade. In contrast, architectures employing skip connections or dense feature reuse such as 50-layer Residual Network (ResNet50), 50-layer Residual Network, version 2 (ResNet50V2), and 121-layer Densely Connected Convolutional Network (DenseNet121), achieve strong accuracy despite declining or fluctuating SNR, indicating that discriminative capability does not solely rely on preserving raw signal strength. More advanced models, including EfficientNet-B0 convolutional neural network (EfficientNetB0) and Inception-ResNet-v2 convolutional neural network (InceptionResNetV2), exhibit heightened sensitivity to corruption, while Extreme Inception convolutional neural network (Xception) demonstrates notable resilience, maintaining high accuracy even under substantial intermediate noise. These findings indicate that SNR alone is not a reliable predictor of classification performance. Instead, robustness emerges from architectural mechanisms that transform, re-weight, and filter features to suppress irrelevant variation while enhancing task-specific information. By linking physical signal characteristics with internal feature representations, this thesis provides new insight into how CNNs manage noisy inputs and offers guidance for designing architectures that remain reliable in the real world, noise prone environments.