{"title":"基于“学习”结构照明的差分衍射网络的高级图像分类","authors":"Jiajun Zhang, Shuyan Zhang, Weijie Shi, Yong Hu, Zheng-Gao Dong, Jiaqi Li* and Weibing Lu, ","doi":"10.1021/acsphotonics.4c0151110.1021/acsphotonics.4c01511","DOIUrl":null,"url":null,"abstract":"<p >As a new optical machine learning framework, the diffractive deep neural network (D<sup>2</sup>NN) has attracted much attention due to its advantages such as low power consumption, parallel computing, and fast execution speed. Here, we demonstrate a new optical neural network design of a differential D<sup>2</sup>NN with structured illumination. In this scheme, the illumination patterns participate in the training process of the network and are optimized by an end-to-end technique. With the application of differential detection, the non-negativity constraint in a diffractive neural network can be alleviated. The test results show that this network architecture can achieve 97.63 and 88.10% classification accuracies on the MNIST and Fashion-MNIST data sets using only one diffractive layer, which exceeds the effect achieved by the five-layer traditional D<sup>2</sup>NN. Moreover, this network architecture can achieve a comprehensive improvement over a traditional D<sup>2</sup>NN in the challenging classification problems of tiny samples and samples blocked by occlusions. Compared with the traditional D<sup>2</sup>NN, this scheme innovatively uses the illumination patterns as new degrees of freedom in system design, which can effectively improve classification ability and reduce the space complexity of the optical neural network.</p>","PeriodicalId":23,"journal":{"name":"ACS Photonics","volume":"11 12","pages":"5289–5298 5289–5298"},"PeriodicalIF":6.7000,"publicationDate":"2024-11-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Advanced Image Classification Using a Differential Diffractive Network with “Learned” Structured Illumination\",\"authors\":\"Jiajun Zhang, Shuyan Zhang, Weijie Shi, Yong Hu, Zheng-Gao Dong, Jiaqi Li* and Weibing Lu, \",\"doi\":\"10.1021/acsphotonics.4c0151110.1021/acsphotonics.4c01511\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p >As a new optical machine learning framework, the diffractive deep neural network (D<sup>2</sup>NN) has attracted much attention due to its advantages such as low power consumption, parallel computing, and fast execution speed. Here, we demonstrate a new optical neural network design of a differential D<sup>2</sup>NN with structured illumination. In this scheme, the illumination patterns participate in the training process of the network and are optimized by an end-to-end technique. With the application of differential detection, the non-negativity constraint in a diffractive neural network can be alleviated. The test results show that this network architecture can achieve 97.63 and 88.10% classification accuracies on the MNIST and Fashion-MNIST data sets using only one diffractive layer, which exceeds the effect achieved by the five-layer traditional D<sup>2</sup>NN. Moreover, this network architecture can achieve a comprehensive improvement over a traditional D<sup>2</sup>NN in the challenging classification problems of tiny samples and samples blocked by occlusions. Compared with the traditional D<sup>2</sup>NN, this scheme innovatively uses the illumination patterns as new degrees of freedom in system design, which can effectively improve classification ability and reduce the space complexity of the optical neural network.</p>\",\"PeriodicalId\":23,\"journal\":{\"name\":\"ACS Photonics\",\"volume\":\"11 12\",\"pages\":\"5289–5298 5289–5298\"},\"PeriodicalIF\":6.7000,\"publicationDate\":\"2024-11-20\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"ACS Photonics\",\"FirstCategoryId\":\"101\",\"ListUrlMain\":\"https://pubs.acs.org/doi/10.1021/acsphotonics.4c01511\",\"RegionNum\":1,\"RegionCategory\":\"物理与天体物理\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"MATERIALS SCIENCE, MULTIDISCIPLINARY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"ACS Photonics","FirstCategoryId":"101","ListUrlMain":"https://pubs.acs.org/doi/10.1021/acsphotonics.4c01511","RegionNum":1,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"MATERIALS SCIENCE, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
摘要
衍射深度神经网络(diffractive deep neural network, D2NN)作为一种新型的光学机器学习框架,以其低功耗、并行计算、执行速度快等优点而备受关注。在这里,我们展示了一种新的具有结构化照明的差分D2NN光学神经网络设计。在该方案中,光照模式参与网络的训练过程,并通过端到端技术进行优化。差分检测的应用可以缓解衍射神经网络的非负性约束。测试结果表明,该网络结构在MNIST和Fashion-MNIST数据集上仅使用一层衍射层就可以达到97.63%和88.10%的分类准确率,超过了传统五层D2NN的分类准确率。此外,该网络架构可以在微小样本和被遮挡的样本具有挑战性的分类问题上实现对传统D2NN的全面改进。与传统的D2NN相比,该方案创新性地将光照模式作为系统设计的新自由度,有效地提高了分类能力,降低了光神经网络的空间复杂度。
Advanced Image Classification Using a Differential Diffractive Network with “Learned” Structured Illumination
As a new optical machine learning framework, the diffractive deep neural network (D2NN) has attracted much attention due to its advantages such as low power consumption, parallel computing, and fast execution speed. Here, we demonstrate a new optical neural network design of a differential D2NN with structured illumination. In this scheme, the illumination patterns participate in the training process of the network and are optimized by an end-to-end technique. With the application of differential detection, the non-negativity constraint in a diffractive neural network can be alleviated. The test results show that this network architecture can achieve 97.63 and 88.10% classification accuracies on the MNIST and Fashion-MNIST data sets using only one diffractive layer, which exceeds the effect achieved by the five-layer traditional D2NN. Moreover, this network architecture can achieve a comprehensive improvement over a traditional D2NN in the challenging classification problems of tiny samples and samples blocked by occlusions. Compared with the traditional D2NN, this scheme innovatively uses the illumination patterns as new degrees of freedom in system design, which can effectively improve classification ability and reduce the space complexity of the optical neural network.
期刊介绍:
Published as soon as accepted and summarized in monthly issues, ACS Photonics will publish Research Articles, Letters, Perspectives, and Reviews, to encompass the full scope of published research in this field.