Research On Flower Recognition Algorithm Based On Convolutional Neural Network

With the rapid development of information technology,the application research of acquiring flower images through mobile devices and being able to quickly and accurately identify and classify flowers has attracted extensive attention.The flower images collected under natural conditions have a large background interference,and because of its own similarity between classes and diversity within the class,it brings greater difficulty in recognition.In view of the challenges faced by flower images recognition,this paper uses a recognition algorithm based on convolutional neural network to identify and classify flowers,and achieves an ideal classification effect.The main contents of this paper are as follows:(1)A flower recognition algorithm with convolutional neural network as the feature extractor is proposed.In this paper,a convolutional neural network with 7convolutional layers,7 batch normalizationl layers,2 maxpooling layers and 1 global average pooling layer is designed as the feature extractor.In addition,Support Vector Machine and Random Forest classifier with better generalization ability are selected to replace Softmax layer of convolutional neural network,and the extracted features are trained and classified,and the classification accuracy of them is compared.Finally,the optimal classification model combining convolutional neural network and Random Forest classifier is obtained.(2)An improved algorithm of feature reuse and feature reconstruction based on MobileNet is proposed.Firstly,the problem of low feature utilization of MobileNet isimproved.By reusing the features of the lower layer,the feature maps of different lower layers are combined into channels and used as the input of the higher layer network,so as to improve the utilization of the feature maps.In this paper,three modes of feature reuse are proposed,and the influence of three modes on the accuracy of flower recognition is analyzed and compared in detail.On this basis,the features extracted by the depthwise separable convolution of the basic structure of MobileNet are reconstructed.Through the loss function to learn the correlation between each feature map,realize features weighting and recalibration of different channels.The learning of effective features is strengthened and useless features is suppressed.The experimental results on the flower data set show that the improved network classification accuracy is higher,it can describe the characteristic information of flowers more comprehensively,and it has less parameter quantity and calculation amount,so it is more suitable to be deployed in the mobile terminal and applied to the field of production and life.(3)At the same time,this paper proposes an improved algorithm based on MobileNetV2,which reuses the feature maps of different layers and reconstruct the features extracted from Inverted residual block.At the same time,the influence of the improved algorithm based on MobileNet and MobileNetV2 on the accuracy of flower classification is compared.