Research On Image Recognition Technology Based On Memristive Neural Network

In the field of image recognition technology,due to the aging of the camera and the complex and changeable external environment,a large amount of noise will be mixed in the collected images,which leads to the low accuracy of image recognition.However,the effective combination of neural network and memristor can not only greatly change the development of artificial intelligence,but also can well suppress the noise contained in the image under the limited dataset.This dissertation first combines the memristor and convolutional neural network for vehicle logo image recognition,and designs a new digital image preprocessing algorithm.In order to simulate the noise in the real scene,four new datasets are constructed based on the VLR-40 dataset;Finally,through two experiments,the image recognition based on memristive neural network is deeply researched and analyzed.The specific work and main contributions are as follows:1.Construct memristive Hopfield neural network based on circuit diagram modeling.In this dissertation,memristor is used to replace the fixed value resistor of the synapse in the traditional Hopfield neural network,which improves the flexibility of the network to deal with different problems and realizes the programming of the synaptic hardware.It provides a theoretical basis for the hardware circuit realization of memristive neural network,and also verifies that memristive crossbar has the characteristic of image storage.2.Construct a memristive crossbar structure for storing grayscale images.This dissertation first constructs a memristive crossbar structure with a size of 224 × 224,and sets the resistance level of the memristor to correspond to the 256-level grayscale of the grayscale image.Then according to the linear mapping relationship,each pixel value of the input gray image corresponds to the input voltage signal of the memristor at each cross point in the memristive crossbar,so as to realize the conversion of the gray image to voltage signals with different amplitudes and the same width.Finally,the weight writing method in offline learning is adopted to realize the storage of 224×224 grayscale images.3.Designed a new digital image preprocessing algorithm.Mean,median and gaussian filtering are commonly used traditional digital image filtering methods.The new digital image preprocessing algorithm designed in this dissertation also has a filtering function,which can suppress the noise contained in the image.4.An image recognition method based on memristive neural network is designed.In the past vehicle logo image recognition algorithms,most of them were time-consuming and did not meet the real-time requirements or the generalization performance of the network model was poor.In this dissertation,a memristive crossbar is constructed based on the characteristics of memristors for the storage of grayscale images,and a new digital image preprocessing algorithm is adopted to achieve data enhancement operations.For the VLR-40 dataset,the recognition accuracy of the memristive neural network is not only higher than that of MobileNetV2,but also can effectively fight the noise in the real scene,which will be of great significance in the field of image recognition.