Design Of Pattern Recognition Model Based On Memristor

In recent years,with the rapid development of pattern recognition technology,it has been widely used in security monitoring,human-computer interaction,medical diagnosis and other fields,bringing intelligent revolution to society.The realization of various algorithms of pattern recognition mainly depends on the transistor based Neumann architecture computer,or even the new non Neumann architecture computing chip.However,due to the separation of "storage" and "calculation" of circuit elements,the problems of "Memory Wall" and "von Neumann Bottleneck" occur,which seriously restrict the operation speed of pattern recognition algorithm and increase the power consumption of the circuit.The development of memristor provides a scientific and feasible way for the integration of information storage and computation.However,at present,the pattern recognition model based on memristor at home and abroad still has some problems such as simple structure,few recognizable categories and low accuracy.In order to solve those above problems,memristor is used in thesis to explore the design of pattern recognition model.Human action recognition is used to test the model.Based on the existing pattern recognition model,thesis proposes a model of multidimensional action signal recognition based on memristor.First,thesis studies the principle of various memristor models and the influence of model parameters on the characteristics of memristor.Then,thesis analyzes and compares the structures and calculation methods of various memristive logic circuits and memristive artificial neural network.And then,thesis discusses the improvement of the performance of power consumption,circuit structure and response speed of the recognition model based on memristor.In thesis,memristor model is built by referring to the parameters of successfully fabricated memristor.For one-dimensional acceleration action signal,the pattern matching model based on memristive logic circuit is designed.For threedimensional video action signal,the memristor cross array is designed and applied to memristive artificial neural network to complete the recognition of multi-dimensional action signal.The main innovations of thesis are as follows:(1)The recognition method of onedimensional action acceleration based on the memristive logic circuit is proposed,which can produce high recognition accuracy.(2)Aiming at the problem that two-dimensional convolution can not protect the integrity of action information of video frame sequence in time dimension,resulting in the low accuracy of video recognition,thesis uses threedimensional memristive convolutional neural network to improve the recognition accuracy of video.(3)Aiming at the problem of low recognition accuracy caused by noise of one-dimensional acceleration action signal collected by acceleration sensor,thesis adopts wavelet denoising method based on threshold to improve the recognition accuracy of acceleration action signal.Experiments show that the proposed pattern recognition model based on memristor can recognize many kinds of actions and has high recognition accuracy.It achieves more than 89% accuracy for one-dimensional acceleration action signal and more than 70%accuracy for three-dimensional UCF101 video action dataset.Thesis has made a useful exploration for the memristor used in new energy-efficient pattern recognition model,and further confirmed the great prospect of memristor in application of pattern recognition.