Robustness Analysis for Deep Learning-Based Image Reconstruction Models

Abstract

Image reconstruction with compressive sensing tra-ditionally employs closed-form and iterative solutions. Deep learning-based image reconstruction methods have emerged as a superior alternative to the classical methods by most mea-sures; they provide enhanced performance in compared metrics. However, there is a lack of studies investigating the robustness of the deep learning-based reconstruction models under different adversaries and comparing them with traditional approaches. Deep learning-based reconstruction models also depend on the measurement matrix they trained with; they perform poorly for other measurement schemes. This behavior does not reflect sparse reconstruction of compressive sensing given a known measurement matrix. That raises the question of whether deep learning-based models learn a general reconstruction function or if they learn a specific relation to a single measurement scenario. In this study, we developed an experiment to analyze the robustness of deep learning-based image reconstruction models under different cases. We used different iterations of a generic deep learning-based image reconstruction model in tests with several techniques used in computer vision to see their effects in the tests. In the end, we presented and discussed the performance of the models.