iii
Abstract
Reconstructing clear and meaningful images from noisy or incomplete data is a funda-
mental challenge in areas such as medical imaging, remote sensing, and computer vision. Tra-
ditional methods like Total Variation and Gauss-Newton often fall short when confronted with
complex shapes or high noise levels, leading to limited accuracy and loss of structural detail.
This dissertation presents a new approach to image reconstruction using a Cascaded Ra-
dial Basis Function Neural Network (CRBFNN). The method features a two-stage neural ar-
chitecture. In the first subnetwork, DBSCAN clustering and K-Nearest Neighbors (KNN) are
used for center and spread estimation, respectively, allowing the model to adapt to the underly-
ing data structure. The second subnetwork applies a fixed spread to ensure stability and com-
putational efficiency during the refinement of the final output. This design enables the network
to respond adaptively to different noise patterns while preserving structural consistency in its
predictions.
Comprehensive experiments were carried out using simulated data affected by white
Gaussian noise, impulsive noise, and contact noise. Across all conditions, the CRBFNN con-
sistently demonstrated strong performance, achieving a Structural Similarity Index (SSIM) of
up to 0.991, a Correlation Coefficient (CC) of 0.983, and a notably low training Mean Squared
Error (MSE) of 0.00066. It also outperformed several modern techniques, including DenseNet,
CWGAN-AM, and Enhanced CNN, both in accuracy and robustness.
Beyond accuracy, the model offers practical advantages. Once trained, CRBFNN pro-
duces high-resolution 2D conductivity maps in approximately 1.3 seconds using only CPU re-
sources, making it suitable for real-time applications and integration into modular EIT systems.
This research highlights CRBFNN as a reliable and efficient tool for image reconstruction under
diverse and challenging conditions. Looking ahead, future work will aim to enhance computa-
tional efficiency through hardware optimization and parallel processing, validate the method
on real-world datasets with complex noise and structural variability, and extend the approach
to 3D and dynamic imaging scenarios. Additionally, integrating CRBFNN with advanced deep
learning architectures such as attention mechanisms, hybrid CNN-RBF models, or perceptual
loss functions may further improve its ability to handle fine structural details and improve gen-
eralization in diverse imaging environments.