Computational ghost imaging based on an untrained neural network

Generally, ghost imaging (GI) experiments involve an object image generated by correlating object and reference beams. A single-pixel detector can detect the interaction between the imaging object and the object beam without requiring any spatial resolution. On the other hand, a spatial resolution detector with no object information is required to detect and record the reference beam. Although the first GI experiment conducted over two decades ago utilized entangled two-photon pairs, subsequent GI experiments have been realized using classical light sources. Based on GI principle, a reference beam can be removed if it can be evaluated numerically. As a result, it is possible to reconstruct the object by computing correlations and even implementing GI into a single-beam structure since the applied random patterns are well known, a method referred to as computational ghost imaging (CGI). Although GI has been applied to numerous fields, expanding its practical applications is hindered by poor imaging efficiency and quality.

Several GI algorithms have been developed to enhance ghost imaging efficiency and quality. Recently, deep learning is increasingly being applied in CGI to solve various problems. It requires thousands of labeled data sets to train a neural network. Acquiring sufficient labeled data for training in GI is always difficult and time-consuming in many practical applications as it uses a single-pixel detector to collect light intensity values. Learning from simulations has been proposed to solve this problem. Although it saves time required to obtain labeled data, it requires a long time to optimize biases and weights of the network and cannot guarantee the validity of the output as the objects are usually far from training data.

To overcome these drawbacks and avoid prior training as well as save training time, a team of researchers from Shandong University: Mr. Shoupei Liu, Professor Xiangfeng Meng, Professor Yongkai Yin, Dr. Huazheng Wu and Dr. Wenjie Jiang proposed a new computational ghost imaging method based on deep learning using an untrained neural network (UNNCGI). The handcrafted network structure was designed by integrating the CGI process into a classical U-net neural network. A 2D object image was reconstructed using a large labeled data set without training. The network input was a set of 1D light intensity values amassed using a single-pixel detector. Furthermore, the neural network was optimized automatically to generate restored images through the interaction between the CGI process and the network. Their research work is currently published in the journal, Optics and Lasers in Engineering.

The authors showed that even without prior training with a large set of labeled data, the untrained neural network could reconstruct the object image by simply inputting a set of 1D light intensity. Compared to the conventional methods, the main advantage of the present method is that it does not require prior training, which saves not only training time but also the time required to train the data. The reconstructed images by UNNCGI exhibited better quality and were much closer to the origin object under different sampling conditions than those of CGI and CSGI. The UNNCGI reconstructed images contained detailed information events at lower sampling ratios.

In a nutshell, a deep learning based CGI method using an untrained neural network was successfully demonstrated. The feasibility of the untrained network was validated by both experiment and simulation. The UNNCGI exhibited superior performance and imaging efficiency than conventional CGI and CSGI. In a statement to Advances in Engineering, Professor Xiangfeng Meng stated that the new reconstruction method of CGI will overcome the inherent limitations of the conventional techniques and promote its practical applications.