Several studies have been conducted in the field of optical computing to deal with the demanding situations of electrical neural networks. Optical computing has many important advantages, such as optical parallelism, which can particularly speed up computing speed, and optical passivity can reduce energy costs and minimize latency. Optical neural networks (ONNs) offer a way to increase computational speed and overcome bandwidth bottlenecks in electrical units. However, ONNs require a coherent laser as a gentle source for calculation and can hardly be combined with a mature device vision formula Soft herbal scenes. Therefore, optoelectronic hybrid neural networks have been proposed, in which the front-end is optical and the back-end is electric. These lens-based formulas increase the difficulty of use in peripheral devices, such as autonomous vehicles.
In a new article in Light: Science
Compared to the hardware architecture in the vision of traditional devices, an optical mask is proposed near the image sensor to update the lenses. According to the theory of geometric optics propagating light in a direct line, scenes can be thought of as sets of points of light. sources, and the optical signal is spatially modulated through the mask to carry out the change and overlap convolution operation on the symbol sensor. It has been verified that the optical mask can update the convolutional layers of neural networks for the extraction of features in the optical domain.
For object classification responsibilities, such as the popularity of handwritten digits, a lightweight real-time popularity network is created to determine the functionality of optical convolution in architecture. Using an unmarried convolution core, the accuracy of popularity can be successful at 93. 47%. the multichannel convolution operation is implemented by having parallel cores in the mask, the sorting accuracy can increase up to 97. 21%. Compared to the vision links of classic devices, it can save around 50% of power consumption.
In addition, by expanding the size of the optical mask, the symbol becomes entangled in the optical domain and the sensor captures a crenellated symbol that is unrecognizable to the human eye, which can encrypt personal data naturally without computational consumption. The functionality of encrypted optics verified in the task of facial popularity. Compared to the random MLS pattern, the popularity accuracy of the jointly optimized mask through an end-to-end network advanced by more than 6%. Along with privacy encryption, it has necessarily achieved the same popularity accuracy functionality. as non-encryption strategies.
Please indicate the appropriate maximum category to facilitate the processing of your application
Thank you for taking the time to provide your feedback to the editors.
Your opinion is for us. However, we do not guarantee individual responses due to the large volume of messages.