Abstract

Strong scattering media bring difficulties to imaging in many fields such as medicine and astronomy. The deep learning method has a powerful fitting ability, which can be better applied in reconstructing the target behind the scattering medium. But the detail of the reconstructed target is often inaccurate. In this paper, the skip connection is added in the neural network to improve the accuracy of the reconstructed detail. This network can combine pixel-level information with high-level semantic information, and the information missed during the encoding process can be more involved in the final reconstruction process. 1100 handwritten characters are used as the targets hidden behind the ground glass. It is found that the quality of the reconstructed target is different when the skip connection is added to the network at different scales. The feature map visualization method is used to help us analyze the role of the skip connection. Meanwhile, PSNR (Peak Signal to Noise Ratio) is also used as an objective evaluation standard to evaluate the quality of reconstructed targets. According to subjective and objective evaluation criteria, conclusions can be drawn that the detail of targets can be better retained when the skip connection is added between the convolutional layer corresponding to the feature map of the size 64*64, and the average PSNR can be enhanced 1 dB compared with the network without skip connections. This work provides a reference for the fusion methods of different scale features in the computational optical imaging.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call