In Computational Spectral imaging, two-dimensional coded apertures and dispersive elements realize the mixed modulation of spatial information and spectral information of the target respectively, and then reconstruct the threedimensional data cube. Therefore, coded aperture plays a vital role. In the imaging process, by moving the coded aperture to increase the number of measurements, the aperture moved one code element at each step to simulate the actual push-broom process. Three types of coded apertures were considered, which are Gauss random coded aperture, Hadamard coded aperture and Harmonic coded aperture, and the reconstruction effect of the three coded apertures were analyzed. The Least Square (LS) algorithm was considered to reconstruct three-dimensional data cube. Compared with the classical Two-step Iterative Shrinkage/Thresholding (TwIST) algorithm, the reconstructed Structural Similarity Index Measurement (SSIM) and Peak Signal to Noise Ratio (PSNR) by LS algorithm were better than TwIST algorithm. It was indicated that the SSIM and PSNR increased with the increasing number of measurements. When the number of measurements was similar with the number of spectral segments, the SSIM of the three coded apertures reached more than 0.9 by LS algorithm. However, the SSIM and PSNR of the Gauss random coded aperture were the largest Obviously, which are 0.995 and 52.560, respectively. And the PSNR of Gauss random coded aperture was 13 dB more than that of Hadamard and Harmonic coded apertures. When the number of measurements was constant, the SSIM and PSNR decrease gradually with the increasing number of spectral segments. The simulation results showed that the LS algorithm was superior to the TwIST algorithm in the reconstruction process, and the Gauss random coded aperture had the best performance.
Super-resolution hyperspectral imaging is a key technology for many applications, especially in the fields of remote sensing, military, agriculture, and geological exploration. Recovering a high resolution image needs enormous data, which puts forward very high requirements on image system hardware. Compressed sampling spectral imaging technology could well solve this problem and achieve high-resolution objects with low-resolution compressed data. In this paper, the method of a compressed sampling spectral imaging based on push-broom coded aperture and dispersion prism is proposed. A spectral aliasing image is formed when the object passing through the dispersive prism. According to the prism dispersion condition and the CCD pixel size, the visible spectrum can be divided into N spectral bands, and the measurement matrix of the coded aperture is respectively calibrated for the center wavelength of each spectral band. By controlling a stepper to implement the push broom of the coded aperture to change the measurement matrix, multiple spectral aliasing images can be obtained. The pixel size of the coded aperture becomes half of the CCD by a relay lens, which means the pixel of CCD is low-resolution for the coded aperture. The super-resolution hyperspectral image of the object is obtained by the improved LS reconstruction algorithm. Simulation results show that, the recovered hyperspectral image has twice resolution compared with the low-resolution CCD image, and the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) increase with the increasing compressed sampling hyperspectral images. For N=31, the average PSNR and SSIM recovered from six aliasing images is 22.019 and 0.235, respectively. The average PSNR and SSIM of the recovered 31 bands are also increasing with increasing aliasing images. While the aliasing imaging is 156, The average PSNR and SSIM exceeds 38 and 0.9. This method proves that super-resolution hyperspectral imaging can be achieved by capturing less low-resolution object images.
In recent years, convolutional neural networks (CNNs) have been widely used in various computer visual recognition tasks and have achieved good results compared with traditional methods. Image classification is one of the basic and important tasks of visual recognition, and the CNN architecture applied to other visual recognition tasks (such as object detection, object localization, and semantic segmentation) is generally derived from the network architecture in image classification. We first summarize the development history of CNNs and then analyze the architecture of various deep CNNs in image classification. Furthermore, not only the innovation of the network architecture is beneficial to the results of image classification, but also the improvement of the network optimization method or training method has improved the result of image classification. We also analyze each of these methods’ effect. The experimental results of various methods are compared. Finally, the development of future CNNs is prospected.