Multi-scale super-resolution reconstruction of terahertz images for postal security inspection

Millimeter-wave and terahertz imaging systems have been used for security checks of human bodies, small postal packages, and industrial nondestructive evaluation. The image resolution is one of the key factors affecting the accuracy of contraband detection and inner defect analysis. We propose a super-resolution reconstruction algorithm based on multi-scale channel attention (MS-CA) residual network to enhance the terahertz images, which are acquired through a homemade real-time terahertz imaging system for postal security inspection. The generalized residual block was introduced and transformed as the basic feature extraction unit, and the channel attention descriptor was inserted in the pooling layer to achieve the channel attention training for the 2D feature map. The fractional matching relationship of pixel coordinates between the HR image and the LR image in the upscale module can achieve reconstruction at any continuous scale. Comparison with traditional deep-learning based algorithms shows that our algorithm outperforms with the highest peak power signal-to-noise ratio (PSNR) and structural similarity index (SSIM) while reconstructing terahertz images at any scale factor.