物理与工程

图像识别是人工智能技术的一个重要应用领域，其中，卷积神经网络在二维图像数据处理方面具有显著优势，是人工智能图像识别的核心技术之一。然而，随着人工智能应用对算力的要求不断提高，传统数字集成电路计算架构下算力瓶颈问题日益凸显。本研究设计了一种基于微透镜阵列的光学卷积系统，将几何光学的空间分割机制与光计算的超高并行特性相结合，在物理层面上完成了卷积乘加运算的映射。系统利用空间光调制器加载图像信息和卷积核，通过微透镜阵列分割光场，利用卷积核对分割后的子区域并行调制，最后经过成像光路汇聚实现求和，从而完成了光学空间上的卷积计算。本研究给出了基于微透镜阵列的光学卷积系统实现原理，设计了实验光路，提出了系统校准、卷积核实现、图像后处理等关键技术的实现方法。实验结果表明，该系统具备基础成像性能，具备实现高斯算子、梯度算子、拉普拉斯算子等复杂特征算子的能力，系统受限于器件精度仍存在一定像差和衍射导致的误差，能够为基于几何光学架构的光学卷积方案在图像预处理任务中的可行性提供验证。

Convolutional neural networks are widely used in tasks such as image recognition due to their excellent performance in two-dimensional data processing.To address the limited AI computing power under the traditional digital integrated circuit computing architecture,this study designs an optical convolution system based on a microlens array.Utilizing the ultra-high-speed and high-parallelism characteristics of optical computing,it achieves convolution computation based on the principles of geometrical optics.The system loads image information and convolution kernels using a spatial light modulator,segments the light field through a microlens array,modulates the segmented sub-regions in parallel using the convolution kernel,and finally converges through the imaging optical path to achieve summation,thuscompletingtheconvolutioncomputationinopticalspace.Thisstudypresentstheimplementationprincipleoftheopticalconvolutionsystembasedonamicrolensarray,designs the experimental optical path,and proposes implementation methods for key technologies such as system calibration,convolution kernel implementation,and image post-processing.Experimental results show that the system possesses basic imaging performance and the ability to implement complex feature operators such as Gaussian operator,Gradient operator,and Laplacian operator.Limited by device accuracy,the system still has errors caused by aberrations and diffraction.It can provide verification for the feasibility of optical convolution schemes based on geometrical optical architecture in image preprocessing tasks.