Adaptive compressed sensing algorithm for terahertz spectral image reconstruction based on residual learning.

[Display omitted] • A new compressed sensing reconstruction method is proposed for fast reconstruction of terahertz spectral images. • Compressed sensing algorithms using deep learning can significantly reduce algorithm complexity and increase reconfiguration speed. • High-quality reconstruction of low-sampling-rate terahertz spectral images. • This work will help expand the cause of terahertz imaging in the field of non-destructive testing. Terahertz time-domain spectroscopy (THz-TDS) is widely applied in the field of rapid nondestructive detection of grain owing to its low photon energy and high penetrating power. Nevertheless, terahertz imaging systems suffer from the problems of long image acquisition time and massive data processing. To mitigate these issues, this work presents an adaptive compressed sensing reconstruction algorithm for terahertz spectral images based on residual learning (ATResCS). The algorithm compresses the number of data samples, reducing the amount of data required for imaging and improving the imaging speed. Further, ATResCS reduces the time complexity by employing a convolutional neural network. The algorithm is validated by acquiring terahertz spectral image data via a THz-TDS system. ATResCS outperforms conventional algorithms regarding peak signal-to-noise ratio (PSNR) and structural similarity, significantly reducing the reconstruction time and, thus, enabling real-time reconstruction. Specifically, at low sampling rates (0.1), ATResCS retains key spectral image information. The average PSNR is 0.96 – 1.015 dB higher than that of DR2-Net, reducing the average reconstruction time by 0.1 – 0.2 s. Experiments demonstrate that ATResCS has better reconfiguration capability and lower algorithm complexity, enabling high-quality and fast reconstruction of terahertz spectral images. [ABSTRACT FROM AUTHOR]