Design And Research Of Three Dimensional Interleaved Photon Convolution Accelerator For Color Image

The development of electronic neural networks is gradually slowing down due to the computational bottlenecks and integration limits of electronic components,as well as the working principles of the von Neumann architecture.As a result,photonic neural networks have been proposed.Based on the characteristics of fast optical computing speed,high parallelism,and low energy consumption,photonic neural networks have to some extent alleviated the shortcomings of electronic neural networks.The photon convolution accelerator implemented through multiplexing technology is suitable for processing high-resolution images and has flexible and adjustable weights.However,this implementation method has certain drawbacks,such as being able to only process single channel images,high spectral resource requirements,and low feature information utilization.In response to these issues,this article proposes solutions and constructs an optoelectronic hybrid neural network combined with an electronic neural network.The main research content of this article is as follows:1.We have designed a photon convolutional accelerator for color images,which utilizes channel shuffling encoding to endow the system with the ability to extract cross channel information while only performing single channel convolution,thus processing color images.Space division multiplexing utilizes limited spectrum resources to efficiently construct multiple convolutional kernels to extract feature information from multiple dimensions and integrate them for utilization.By combining electronic neural networks,an optoelectronic hybrid neural network is constructed to achieve classification and prediction of high-resolution color images.The channel shuffling operation improved the accuracy of network classification by 0.87%,reaching a maximum of 97.36%;By using multi-dimensional feature extraction and integration to process grayscale images,the accuracy was improved by up to 1.49%,and the processing of channel shuffle images was improved by up to 1.11%;The combination of the two operations results in a classification accuracy of up to 97.90%.2.According to the design scheme,the channel shuffling operation is completed during image data encoding,and the power division function of the waveform shaper is used to achieve spatial division multiplexing of the spectrum and efficient construction of convolutional kernels,thereby completing multi-dimensional feature extraction,which is then received,restored,and integrated by the computer.The specific construction of the photon convolution accelerator was completed in the experimental environment,and the construction process of the photon convolution kernel and the calculation process of photon convolution were recorded and displayed.Afterwards,the electronic neural network was trained based on simulation,and the overall construction of PCACI-ONN was ultimately achieved.For some cat and dog datasets,multiple photon convolution treatments were attempted to analyze their convolutional effects.Finally,the classification performance of the network was tested,and its accuracy was greater than or equal to 95%.