Terahertz (THz) radiation is electromagnetic radiation of 0.1 to 10 THz. It lies between radio waves and light waves, millimeter waves and infrared in frequency, and between electrons and photons in energy. On both sides of the electromagnetic spectrum, infrared and microwave technologies are well established, but THz technology is still largely a gap. The reason for this is that in this band, it is neither entirely suitable for treatment by optical theory nor for microwave theory to be studied. THz systems have a wide range of applications in semiconductor materials, the study of the properties of high-temperature superconducting materials, tomographic techniques, unlabeled genetic examinations, imaging at the cellular level, chemical and biological examinations, as well as broadband communications, microwave directional and many other fields. However, the degradation of THz images caused by diffraction phenomena and system noise is still an urgent problem. In 1972, Richardson used an iterative Bayesian-based method to reduce PSF diffraction degradation, which recovers clean images through deconvolution [1]. In [2], By quantizing based on the idea of maximum entropy, the author was able to produce clear images and distributions from noisy and incomplete data. In recent years the number of methods used to incorporate terahertz image recovery has increased due to the great role of neural network feature extraction [3], [4], [5], [6]. Furthermore, the success of the single THz image recovery has also been extended to THz image super-resolution, THz image detection, and other research areas.

The reasonableness of the generated image depends on the synthesis method and model used, as well as its alignment with the actual imaging system. If the PSF function model accurately captures the characteristics of the imaging system and considers various influencing factors during synthesis (such as signal attenuation, scattering, absorption, etc.), then the generated image may be quite reasonable. In this paper, the PSF is the main factor causing the blurring of terahertz images. We use the AI approach to solve THz image recovery in the traditional methods. The real THz image obtained from the scan is of low resolution due to the limitations of the photographic elements and cannot be used for network training due to the lack of labels. The objective is to overcome the diffusion effect caused by the THz PSF, different from other image restoration tasks. For example, Fig. 1 shows several existing methods for recovering THz images. It can be seen that the existing methods are difficult to solve the problem.

In the region where the luminance changes, the image restoration quality would be relatively low. The Unet method in Fig. 1 does not completely remove THz diffraction and it is difficult to extract low-level feature maps. Based on the resnet structure [7], which does not correctly capture the background color of the original image and use it to recover the THz image.

Although THz technology is widely used in military, industrial and agricultural applications, there are still challenges in the following areas of imaging technology.

Most of the existing methods use the frequency characteristics of terahertz waves reflected by the target, or the THz images to analyze the target information. This capture method is direct measurement, which does not consider the PSF effect. As the sampling images are seriously affected by the PSF, it is necessary to adopt the PSF to improve the imaging quality.

There is rarely literature on the use of real PSF to synthesize THz images. Moreover, the imaging equipment is expensive and hard to obtain, which is not favorable for general researchers to work in this field.

Traditional methods such as deconvolution are very ineffective in recovering PSF, making it difficult to extract meaningful background images.

Therefore it is necessary to establish a complete system to solve the problem. To deal with these problems, we build a THz imaging system and propose a Depthwise Dense Instantiation Normalization Network (DwDINet) to overcome the diffusion effect caused by the THz PSF. The motivation behind the proposed method is that the multi-scale nature of THz images can guide image recovery. Fig. 5 shows the overall architecture of the network. To make the network more lightweight, we use Depthwise separable convolution [8] and ConvNeXt [9]. Moreover, inspired by HINet [10], we use IN to further design our DwDIN module. And we use Full-connected Composition [11] to implement the connection method between DwDINs. The main contributions of the paper are as follows:

We obtained the PSF synthetic THz dataset by using the built THz imaging system and made it publicly available to researchers;

Based on the DwDIN block, we propose the DwDINet network to solve the THz image recovery problem. Also, our method achieves state-of-the-art results compared to other methods;

We performed extensive experiments on other low-level image recovery tasks and demonstrated that the proposed DwDINet network is important for the field of image recovery.