Accelerated Blind Denoising of GPR Data via Deep Random Projections

Ground Penetrating Radar (GPR) represents a critical non-invasive technology for subsurface imaging across diverse applications including civil infrastructure assessment, archaeological surveys, and geological exploration. However, GPR data quality is inherently compromised by multiple noise sources including electromagnetic interference, thermal noise, surface clutter, and system-related artifacts, which collectively obscure important subsurface features and complicate interpretation. This paper presents an application of Deep Random Projections (DRP), a computationally efficient variant of the Deep Image Prior framework, for blind denoising of real-world GPR data where neither noise characteristics nor ground truth clean signals are available. Our approach leverages the implicit regularization provided by convolutional neural network architectures while dramatically reducing computational requirements by freezing network weights and optimizing only the input seed. Extensive experiments on field-collected GPR data demonstrate that DRP achieves denoising performance comparable or superior to state-of-the-art methods including S2S-WTV while requiring two orders of magnitude fewer trainable parameters and achieving 5-10× speedup in processing time. The method's ability to operate without training data or explicit assumptions about noise distributions makes it particularly suitable for practical GPR applications where obtaining clean reference data is infeasible.