Accelerating ML-based super-resolution for gigapixel-scale acoustic imaging

Abstract Gigapixel-scale acoustic imaging is increasingly used in biology, materials science or industrial failure-analysis to capture fine structural details across large fields of view (FOV). Automated analysis is essential for interpreting these datasets, yet a large pixel-count does not guarantee adequate image quality for downstream analysis. Machine learning (ML)-based image enhancement, including super-resolution (SR), can address this limitation, but inference times and memory demands grow substantially with increasing image size. Here, we exploit strategies for improving the efficiency of SR models for scanning acoustic microscopy, where automated analysis of large FOV gigapixel images is essential. By combining insights from neural scaling laws with architectural and runtime optimizations, we are able to reduce evaluation time and memory footprint by roughly an order of magnitude while maintaining reconstruction quality. Beyond acoustic applications, the presented optimization strategies may also be applicable to ML-based SR workflows in other gigapixel imaging domains.