.png){kind=icon}
.png){kind=icon}
Return
Height quantized diffractive deep neural networks
.png){kind=icon}
PRE
AI
DOI：10.1088/1402-4896/adb651.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diffractive deep neural networks (D2NN) have shown advantages in artificial intelligence image processing tasks, such as lens-free imaging. Since the accuracy and resolution of current fabrication technologies such as photolithography and 3D printing are difficult to meet the low-cost D2NN manufacturing, in order to solve this problem, we designed height quantized diffractive deep neural networks, thereby improving the networks detail resolution and reducing the fabrication accuracy requirements. We experimentally verified the functionality of the proposed networks, and simulation results show that this structure can achieve the same training effect with less training time. In addition, the quantization process is introduced into D2NN as a kind of noise, which can partially avoid the overfitting of the D2NN.
Keywords:
diffractive deep neural networks
computational imaging
diffractive imaging
Journal
IF:
2.6
Papers:
4.3K
Citations:
2.5W