THz time-domain spectral imaging enhancement based on an MLP-SRCNN composite network

Abstract

A enhancement algorithm for scanning images from terahertz time domain spectral system is proposed. The algorithm is a machine learning algorithm based on a multi-layer perceptron (MLP) network and a super-resolution convolutional neural network (SRCNN). The MLP network is utilized to convert a time domain waveform vector into a scalar pixel intensity adaptively; the SRCNN network aims to achieve the spatial enhancement of the output from the MLP network. The new algorithm needs only two scanning images to achieve the enhancement effect. The two images are captured before and after a rigid body translation of the specimen. Then, a classic relative deformation algorithm, digital image correlation (DIC), is used to compute the displacement fields between the two images. The difference between the calculated and actual displacement fields is taken as the loss function of the MLP-SRCNN composite network. The effectiveness of the proposed algorithm is demonstrated by a typical validation experiment. A considerable enhancement on the quality of the scanning images from THz-TDS is achieved under the no need of additional training data.