Research On Optical Diffraction Neural Network For Target Recognition

As one of the most important technologies in the 21st century,artificial intelligence is influencing and changing the society with unprecedented depth and breadth.The realization of artificial intelligence greatly depends on artificial neural network.However,the slowdown of Moore’s law in the "post Moore era" and the tidal data load problem caused by the "von Neumann bottleneck" make the von Neumann computing system based on silicon-based electronic chip unable to meet the growing demand for large-scale data processing and computing power of artificial neural network due to huge energy and time consumption.The optical diffraction neural network constructed with light as the operation carrier of artificial neural network has become one of the most potential solutions to break through the above technical bottleneck because of the information transmission characteristics of parallel light speed,low energy consumption and low time consumption.However,there are still the following problems in the research field of optical diffraction neural network: on the one hand,facing the system construction theory,the design theory of optical diffraction neural network lacks quantitative analysis means,which leads to the contradiction between system design and processing,and limits the design and processing realization of the system in the visible band;On the other hand,for the system application scenario,the deception attack caused by the disadvantages of open-loop design of optical diffraction neural network and the application scenario limitation caused by the limitation of single wavelength optical diffraction neural network design method under multi wavelength function restrict the practical application of the system.In this paper,"Research on optical diffraction neural network for target recognition",theoretical analysis and experimental research are carried out to solve the above problems.The research results of this paper may provide reference for improving the theoretical system of optical diffraction neural network and enriching the application scenarios of optical diffraction neural network.The main research contents of this paper include:1.Aiming at the contradiction between the design and fabrication of visible light band system caused by the traditional maximum half cone diffraction angle theory,the full connection model of visible light optical diffraction neural network is studied.It is clear that the system design should be based on the full connection topology in the framework of artificial neural network.Through the quantitative analysis of the restrictive relationship between the physical parameters required in the construction process of the system,the formula of the maximum half cone diffraction angle is revised and the full connection model is improved.The proposed full connection model based on the revised formula of the maximum half cone diffraction angle defines the quantitative description method and fine regulation means of the physical parameters required by the system to realize the full connection topology.The system parameters are flexibly designed according to the model,which effectively avoids the processing problems caused by the traditional maximum half cone diffraction angle design theory under visible light.The above research provides a theoretical basis and parameter selection guidance for the design and preparation of visible light optical diffraction neural network in this paper.2.Aiming at the serious coherent noise of single wavelength optical diffraction neural network system,the multi wavelength optical diffraction neural network is studied.A design scheme of multi wavelength and incoherent optical diffraction neural network system is proposed based on the excitation function model constructed by spatial partition of detection area.Theoretical analysis shows that the designed system can effectively suppress coherent noise because the output intensity is the sum of incoherent narrowband light intensity.Taking red light(700 nm),green light(546.1 nm)and blue light(435.8 nm)as incident light sources,based on the target recognition task of handwritten numerals 0-9,the multi wavelength response and incoherent light response of the system are simulated and verified.Simulation analysis shows that the proposed excitation function model improves the numerical classification accuracy of the system in the task of target recognition;The numerical classification accuracy better than 90% at each wavelength means that the system design method can obtain good performance in target recognition task;The designed system can correctly realize the target recognition task of handwritten digits under at least ten wavelengths.The experimental system is built with two wavelengths of lasers(637 nm and 561 nm).The experiment results verify the correctness of the theoretical design.3.Aiming at the problem that optical diffraction neural network can not resist spoofing attack,the target recognition task of visible light optical diffraction neural network is studied.Based on the target recognition task of 0-9 handwritten numeral samples,the deception attack caused by the disadvantages of the open-loop design of the system is clarified: that is,because there are only detection areas one-to-one corresponding to the tested target category,the system will still focus the maximum light intensity in a certain area for the deception target,and then misjudge the deception target as one of the original ten categories of samples.By making the deception target data sample set and reserving additional detection area,the closedloop design idea of the system is proposed.On the basis of retaining the very similar function of the system to the original target,the closed-loop classification function of judging many common types of deception targets(occlusion / alteration samples)as deception samples is realized.The experimental system is built based on 632.8 nm He-Ne laser,and the consistency between the experimental results and the theoretical design is verified.