Non-iterative Deep Neural Networks for Compressed Coded Aperture Imaging
Zhang Cheng,Chen Mingsheng,Chen Qianwen,Shen Chuan,Wei Sui
DOI: https://doi.org/10.1007/s12596-022-01033-5
2023-01-01
Abstract:Compressed Imaging (CI) provides an optical implementation of Compressed Sensing which replaces the conventional imaging systems with the new design for the aim of reducing the sensor cost, acquisition time and so on. However, there are several challenges existed in CI implementation, which involved a large amount of data, imposed significant limitations on optical system design, calibration, data storage requirements, and computational complexity, etc. Compressed Coded Aperture Imaging (CCAI) is a practical compressed imaging system, which can reconstruct the true image from a single, low-resolution, noisy coded measurements. However, the iterative scheme of the traditional reconstruction algorithm makes the algorithm costly, and the computational complexity is high, and thus real-time reconstruction becomes more challenging. To address the challenges raised, this paper introduces deep learning into the reconstruction process of CCAI. The training data pairs composed of coded measurements and original images are used to train the deep learning architecture, and the reconstruction process is completed by non-iteratively, real-time reconstruction from low-resolution coded measurements. Furthermore, in the training phase, the singular value decomposition based preprocessing module is added to eliminate the correlation between each measurement and thus reduce the mutual coherence of the forward operator in CCAI, which can greatly enhance the reconstruction performance. Numerical experiments demonstrate that our proposed method can achieve high-quality and real-time reconstruction.