Miniaturized Spectroscopy Based on Lens-Free Speckle Image Encoding

Abstract

Miniaturized spectroscopy has attracted extensive interest recently due to its advantages of high integration and low cost. Image encoding is an emerging and promising technique to connect in-plane field distributions and spectral components, considering its perfect match with mature image sensors in portable platforms. Usually, an object lens and a large working distance are required to resolve the wavelength-dependent images encoded by light dispersion nanostructures, which inevitably increase the complexity for optical alignment and the system footprint. Here, a lens-free speckle image encoding technique was developed for computational spectroscopy by integrating disordered gold nanorods onto a CMOS image sensor with an entire thickness of less than 2 mm. A convolutional neural network (CNN) algorithm was applied to decode the image-spectrum relation by classifying the images with a multilayer perceptron to extract high-level features, overcoming the high correlations of the encoded images in such a compact configuration. Accurate spectral reconstruction was demonstrated with a peak wavelength deviation of less than 1 nm. Furthermore, the required CNN training data set scale or training time was found to be greatly reduced by simply embedding metallic micro holes between the nanorods and the image sensor due to the remarkably improved image quality via the pinhole imaging principle. It is expected that such a compact and low-cost miniaturized spectroscopy platform with decent spectral sensing accuracy enhanced by the deep-learning technique demonstrates important application prospects for on-site inspection and a distributed sensor network.