Research Of Non-imaging Object Classification Based On Machine Learning

With the advancement of technology,object classification has received more and more attention,and it could be indispensable in many fields,such as military reconnaissance,medical diagnosis,video surveillance and so on.In recent years,a strategy of skipping the imaging step and directly classifying the object after detection has been proposed.The scholars have proposed some object classification schemes.However,the current nonimaging object classification schemes cannot perform under photon-limited conditions.With the rapid development of photon detectors and imaging systems,a single-photon detector-based correlation imaging system can solve the problem of photon-limited imaging,and because of its unique imaging mechanism that separates data acquisition from imaging,the system is employed as the data acquisition platform.A non-imaging object classification scheme is proposed.The research in this paper mainly includes:1.Research on non-imaging object classification based on machine learning.In the field of correlation imaging,the Hadamard matrix has been widely concerned due to its high imaging efficiency.A scheme by using Hadamard modulation is proposed to achieve the goal of achieving efficient object classification.We discuss the construction of the system,the acquisition of data,the construction of machine learning algorithms,and the selection of data sets.Then the simulated experiment is carried out to verify the feasibility of the scheme,and the classification accuracy under different photon number levels is further discussed.The bottleneck of photon-counting method for classification under extremely low-light conditions is analyzed.2.Research on non-imaging object classification under extremely lowlight conditions based on machine learning.Since the photon-counting method is very sensitive to the photon level,such an algorithm cannot achieve an ideal object classification under photon-limited conditions.Therefore,a scheme based on the first-photon pulse count is proposed.For this program,we conduct simulation experiments and actual experiments.In the actual experiment,for the problem that it is difficult to obtain a large number of training sets,a strategy of generating training sets by simulation is proposed.The experimental results show that our scheme effectively realizes object classification under extremely low-light conditions,and still has an accuracy of up to 90% in the case where images cannnot be constructed by imaging algorithms.This paper focuses on solving the problem of object classification under extremely low-light conditions.From the introduction of the program,to the data acquisition and processing,from simulation experiments to actual experiments,the object classification scheme is introduced and analyzed.We believe that this solution can provide a new solution to object classification in some extreme environments.