image: a, Sketch of the apparatus. Speckle illumination modes generated by the RGG were divided into a reference path that was directly measured by a pixelated camera and an object path that was measured by a single-pixel detector after interacting with the sample. b, Raw speckle patterns H (top) and intensity sequence I (bottom) measured by the camera and the single-pixel detector, respectively. c, Correlating H and I one can get a low-quality result especially when the sampling ratio β in our case is as low as 10%. Then, we feed it into the neural network and keep it fixed. The output of the neural network is taken as the estimated object, which is then numerically multiplied with H to generate the estimated intensity sequence. We measure the mean square error between I and the estimated intensity sequence as the loss function to guide the update of weights in the neural network. view more
Credit: by Fei Wang, Chenglong Wang, Mingliang Chen, Wenlin Gong, Yu Zhang, Shensheng Han, and Guohai Situ
In microscopy imaging, the super-resolution techniques that based on frequency shift can increase the resolution by 2 times. In fluorescence microscopy, such as STORM, STED and other technologies can increase the spatial resolution to more than 10 times the diffraction limit, and the latter won the 2014 Nobel Prize in Chemistry. However, these technologies are difficult to directly promote and apply to long-distance super-resolution imaging. Increasing the aperture of the imaging system and shortening its focal length are the major way to improve the resolution of long-distance imaging system, which is still restricted by the diffraction limit.
In a new paper published in Light Science & Application, a team of scientists, led by Professor Guohai Situ from Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, China, and co-workers have developed an AI-driven super-resolution technique. First, they use GI, a technology with high information acquisition efficiency and high detection sensitivity, to encode the object information into single-pixel measurements. Second, they restore the object through a physics-enhanced deep neural network (DNN). Specifically, they feed the low quality GI image obtained by linear correlation to the DNN, take the output of the DNN as the estimation of high quality GI image, and calculate the estimation of GI measurements starting from the DNN output by the forward propagation model of GI. Then, the parameters in the DNN were updated to minimize the error between the raw and estimated GI measurements. Along with the minimization of error, the output of DNN also converges to a much better result. They call the proposed method Ghost Imaging using Deep neural network Constraint (GIDC).
“GIDC does not need lots of labeled data pairs to train a DNN, all it needs is the raw GI measurements and the forward propagation model of GI both are available in a typical GI system, said by Guohai Situ. Its actually a general method which does not bias towards any dataset.
They demonstrated their methods on a pseudothermal GI system. In which, the random illumination speckle field is divided into a reference path and a test path. The light of the reference path is directly detected by a camera without interacting with the object, and the light of the test path is received by a single-pixel detector after passing through the object. Although neither detector directly records a resolvable image of the object, one can employ an intuitive linear algorithm to reconstruct its image by spatial correlating the acquired time-varying patterns and the synchronized bucket signal.
The researchers performed a comparative study on the base of a number of challenging real-world scenarios including a flying drone, and demonstrate that the proposed method outperforms other widespread GI methods.
“GIDC has the potential to break the diffraction limit of GI and reduce the sampling ratio required to obtain a high signal-to-noise ratio (SNR) image., said Situ.
“Compared with conventional imaging, GI can collect more information about the object, but it is highly coupled to the 1D bucket signal. The proposed method allows more information to be effectively extracted and is suitable for many computational imaging systems that adopt encoding-decoding strategy, they added.
“Considering that the GI is a promising method for long-distance imaging, we believe the proposed method has great potential for high resolution remote sensing. the scientists forecast.
Journal
Light Science & Applications