Single shot real-time high-resolution imaging through dynamic turbid media based on deep learning

•We scrutinize the Fourier-domain shower-curtain effect (FDSE)-related noise model, which makes it possible to simulate adequate training data to learn FDSE correlography problems.•Without knowing the experimental scene, the generated convolutional neural network (CNN) is computationally efficient and extremely robust to various forms of noise, far exceeding the capabilities of existing algorithms.•The targets can be reconstructed from a standard sCMOS detector with a 150 ms exposure. Optical experiments have proved that in dynamic turbid media, the real-time high-resolution imaging with a single lens is possible. Low signal-to-noise ratio (SNR) measurement is conceivable the primary obstruction to real-time, high-resolution through dynamic turbid media optical imaging. To break this restriction, by individualizing and employing these low SNR measurement data, the spectrum estimation theory is procured a noise model for scatter imaging. The noise model proposed is exploited to synthesize data set training to settle the related problems of noise phase without knowing the experimental scenes. We verify the robustness of the resulting deep correlography method to noise, outdistance the capabilities of the existing Fourier-domain shower-curtain effect (FDSE) system in terms of spatial resolution and total acquisition time, in addition, the targets can be reconstructed from a standard sCMOS detector with a 150 ms exposure.