Color computational ghost imaging by deep learning based on simulation data training

Abstract

We present a new color computational ghost imaging strategy using a sole single-pixel detector and training by simulated dataset, which can eliminate the actual workload of acquiring experimental training datasets and reduce the sampling times for imaging experiments. First, the relative responsibility of the color computational ghost imaging device to different color channels is experimentally detected, and then enough data sets are simulated for training the neural network based on the response value. Because the simulation process is much simpler than the actual experiment, and the training set can be almost unlimited, the trained network model has good generalization. In the experiment with a sampling rate of only 4.1%, the trained neural network model can still recover the image information from the blurry ghost image, correct the color distortion of the image, and get a better reconstruction result. In addition, with the increase in the sampling rate, the details and color characteristics of the reconstruction result become better and better. Feasibility and stability of the proposed method have been verified by the reconstruction results of the trained network model on the color objects of different complexities.

Color ghost imaging through a dynamic scattering medium based on deep learning

“…Compared with the stable condition described in Ref. 13, the undulation caused by the randomness and uncertainty of dynamic scattering medium is inevitable, and this serious vulnerability may lead to poor performance or even failure of the trained network model.…”

“…n means the noise introduced by the scattering medium and experiment conditions, and B is the detected light intensity sequences (LIS). Referencing the linear intensity correlation reconstruction methods in ghost imaging, 13 an object image is obtained as E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 2 ; 1 1 6 ; 4 8 8…”

“…n means the noise introduced by the scattering medium and experiment conditions, and B is the detected light intensity sequences (LIS). Referencing the linear intensity correlation reconstruction methods in ghost imaging, 13 an object image is obtained as OjGI=1M(Sj−I⟨Sj⟩)T(B−I⟨B⟩)=1MΨjT(krΨrOr+kgΨgOg+kbΨbOb)=krAjrOr+kgAjgOg+kbAjbOb,where …”

“… – 12 The feasibility of this method has been verified in binary computing ghost imaging through a static scattering medium. Another work that we recently reported uses simulated data sets rather than the actual one to reduce the workload and enhance the CCGI scheme, 13 but when applying this simple and effective method to CCGI through dynamic scattering medium, a fatal flaw is exposed: unlike the training set consisting of experimental data, the simulation data cannot reflect the non-ideal characteristics brought by the experimental devices and environment. Compared with the stable condition described in Ref.…”

VGenNet: Variable Generative Prior Enhanced Single Pixel Imaging

“…The generative adversarial network (GAN) has also been used for SPI, in which the discriminator introduces adversarial error into the loss function. The GAN-based method is able to achieve better results due to the advanced adversarial training strategy. − Other studies are mainly carried out from the aspects of learning method, , network structure, , and the combination with specific applications. − …”