基于 DD-CASSI 退化成像模型的高光谱图像重建算法(内封面文章·特邀)

Objective Coded Aperture Snapshot Spectral Imaging (CASSI) is a typical computational optical imaging technique which is based on the principle of compressed sensing to restore spectral and two-dimensional raw data. After the detection of high-dimensional data, the quality of the reconstruction algorithm determines the detection performance of the CASSI system. However, current reconstruction algorithms are still facing some challenges. One of the most significant challenges is that many reconstruction algorithms assume that the CASSI system employs the idealized sampling measurement model, failing to adequately account for critical factors such as instrument calibration and spectral transmittance correction. This limits the application of these algorithms in real imaging systems. Therefore, it's necessary to further optimize and refine the algorithm to enhance its stability and adaptability in practical applications. For this purpose, an optical degraded imaging model based on DD-CASSI system and a scheme for applying the degraded imaging model to optimize image reconstruction algorithms is designed. Methods Based on the coded aperture compression imaging system with dual grating dispersion structure, an optical degraded imaging model incorporating the system point spread function (Fig.1) is constructed. An optimization scheme is proposed that applies the degraded imaging model to traditional hyperspectral image reconstruction methods and reconstruction methods based on end-to-end deep learning network (Fig.3). Furthermore, a deep unrolling network structure combined with the degraded model (Fig.5) was designed. The architecture integrates a Half-Quadratic Splitting algorithm, convolutional neural network and multi-head attention modules (Fig.7). In comparison with alternative algorithms, this network has been demonstrated to reconstruct spectral curves with greater smoothness and higher correlation to the reference spectral curve. Results and Discussions When the degraded imaging model is applied to iterative algorithm optimized by the traditional regular term (Fig.2), this model attains a Peak Signal to Noise Ratio (PSNR) enhancement that exceeds 5 dB for targets with simple spatial structures and spectral characteristics. For more intricate target scenarios, this enhancement diminishes to 2-3 dB. It also improves the Spectral Angle Mapper (SAM) metrics by approximately 0.14 rad. When applied to end-to-end deep learning networks (Fig.4), training the network on datasets generated by introducing a degraded imaging model has been shown to achieve an improvement in PSNR of over 10 dB and in SAM metrics of over 0.1 rad. Furthermore, the constructed deep unrolling network structure attains a PSNR enhancement of approximately 10 dB and a SAM enhancement of 0.05 rad in comparison with traditional iterative algorithms. A comparison of PSNR with deep learning networks demonstrated an enhancement of approximately 3 dB, while SAM exhibited an improvement of 0.02 rad. The proposed model also demonstrates significant advantages over λ-Net in terms of the number of model parameters and reconstruction time (Tab.3). Conclusions The effectiveness of the degenerate imaging model when applied to the task of hyperspectral image reconstruction provides an efficient solution for the reconstruction process of high-resolution hyperspectral images. the degraded imaging model and reconstruction algorithm system in enhancing the reconstruction quality and system performance of digital-mirror-device-based encoded aperture snapshot spectral polarization images and provide theoretical support for further optimization of the design and high-precision reconstruction method of snapshot spectral polarization imaging system. The research results are of great significance to enhance the application value of digital-mirror-device-based coded aperture snapshot spectral polarization technology in the fields of dynamic target detection and multidimensional information fusion imaging.