RESUMO
Compressive imaging allows one to sample an image below the Nyquist rate yet still accurately recover it from the measurements by solving an L1 optimization problem. The L1 solvers, however, are iterative and can require significant time to reconstruct the original signal. Intuitively, the reconstruction time can be reduced by reconstructing fewer total pixels. The human eye reduces the total amount of data it processes by having a spatially varying resolution, a method called foveation. In this work, we use foveation to achieve a 4x improvement in L1 compressive sensing reconstruction speed for hyperspectral images and video. Unlike previous works, the presented technique allows the high-resolution region to be placed anywhere in the scene after the subsampled measurements have been acquired, has no moving parts, and is entirely non-adaptive.
RESUMO
Computer vision with a single-pixel camera is currently limited by a trade-off between reconstruction capability and image classification accuracy. If random projections are used to sample the scene, then reconstruction is possible but classification accuracy suffers, especially in cases with significant background signal. If data-driven projections are used, then classification accuracy improves and the effect of the background is diminished, but image recovery is not possible. Here, we employ a shallow neural network to nonlinearly convert from measurements acquired with random patterns to measurements acquired with data-driven patterns. The results demonstrate that this improves classification accuracy while still allowing for full reconstruction.
RESUMO
Both 3D imaging and hyperspectral imaging provide important information of the scene and combining them is beneficial in helping us perceive and understand real-world structures. Previous hyperspectral 3D imaging systems typically require a hyperspectral imaging system as the detector suffers from complicated hardware design, high cost, and high acquisition and reconstruction time. Here, we report a low-cost, high-frame rate, simple-design, and compact hyperspectral stripe projector (HSP) system based on a single digital micro-mirror device, capable of producing hyperspectral patterns where each row of pixels has an independently programmable spectrum. We demonstrate two example applications using the HSP via hyperspectral structured illumination: hyperspectral 3D surface imaging and spectrum-dependent hyperspectral compressive imaging of volume density of participating medium. The hyperspectral patterns simultaneously encode the 3D spatial and spectral information of the target, requiring only a grayscale sensor as the detector. The reported HSP and its applications provide a solution for combining structured illumination techniques with hyperspectral imaging in a simple, efficient, and low-cost manner. The work presented here represents a novel structured illumination technique that provides the basis and inspiration of future variations of hardware systems and software encoding schemes.