Light-field detection measures both the intensity of light rays and their precise direction in free space. However, current light-field detection techniques either require complex microlens arrays or are limited to the ultraviolet–visible light wavelength ranges^1–4. Here we present a robust, scalable method based on lithographically patterned perovskite nanocrystal arrays that can be used to determine radiation vectors from X-rays to visible light (0.002–550 nm). With these multicolour nanocrystal arrays, light rays from specific directions can be converted into pixelated colour outputs with an angular resolution of 0.0018°. We find that three-dimensional light-field detection and spatial positioning of light sources are possible by modifying nanocrystal arrays with specific orientations. We also demonstrate three-dimensional object imaging and visible light and X-ray phase-contrast imaging by combining pixelated nanocrystal arrays with a colour charge-coupled device. The ability to detect light direction beyond optical wavelengths through colour-contrast encoding could enable new applications, for example, in three-dimensional phase-contrast imaging, robotics, virtual reality, tomographic biological imaging and satellite autonomous navigation.

MEMS (Micro Electro Mechanical System) gyroscopes have been widely applied to various fields, but MEMS gyroscope random drift has nonlinear and non-stationary characteristics. It has attracted much attention to model and compensate the random drift because it can improve the precision of inertial devices. This paper has proposed to use wavelet filtering to reduce noise in the original data of MEMS gyroscopes, then reconstruct the random drift data with PSR (phase space reconstruction), and establish the model for the reconstructed data by LSSVM (least squares support vector machine), of which the parameters were optimized using CPSO (chaotic particle swarm optimization). Comparing the effect of modeling the MEMS gyroscope random drift with BP-ANN (back propagation artificial neural network) and the proposed method, the results showed that the latter had a better prediction accuracy. Using the compensation of three groups of MEMS gyroscope random drift data, the standard deviation of three groups of experimental data dropped from 0.00354°/s, 0.00412°/s, and 0.00328°/s to 0.00065°/s, 0.00072°/s and 0.00061°/s, respectively, which demonstrated that the proposed method can reduce the influence of MEMS gyroscope random drift and verified the effectiveness of this method for modeling MEMS gyroscope random drift.