Application Of Convolutional Neural Network In Image Classification And Recognition

Image classification is one of the important research directions in the field of computer vision.Based on the existing image classification methods,this paper focuses on the application of convolutional neural networks in different classification tasks.First,through the reproduction of the classical model and comparison experiments on different data sets,we summarized how to design a good classification model.Then combined with the actual data,three improved convolutional neural network classification models are proposed and good results are obtained in different classification tasks.The main contributions of this article are as follows:First,three models of Alex Net GoogLeNet Res Net are implemented with the help of the MXNET deep learning framework.The effects of different models on classification are analyzed through a comparative experiment on the Fashin-MNIST,Cifar-10 dataset,and the methods of how to optimize the design model from parameter and structure are summarized.Secondly,an improved convolution neural network model is proposed.For the license plate digital character data set with high noise,the BatchNorm algorithm is fused between the convolution layer and the activation layer,and the Dropout algorithm is introduced into the full connection layer.The robustness and convergence of the model are increased.The experimental results show that the fusion can deal with high noise data effectively.In the case of low noise,moderate noise and strong noise pollution,our method has obvious advantages in digit character recognition,compared with BP,PCA and so on.The average recognition rate is nearly 5% higher than that of other methods,especially in the case of strong noise pollution data,and the recognition rate is nearly 20% higher than that of template matching method.Thirdly,an improved siamese-convolution neural network algorithm is proposed for face recognition,which is composed of two identical convolution neural networks and shares network weights.The convolution structure effectively removes the external noise interference.In the training of the structure,a learning algorithm of depth difference measurement is adopted.In nonlinear dimensionality reduction,the weight sharing structure can automatically extract the same features and the DDML algorithm increases the effectiveness of feature extraction.The experimental results on ORL,Yale B and AR face databases show that the recognition stability is higher than that of PCA,CNN,and the recognition rate is increased by 5%.Fourthly,an improved full convolution neural network model is proposed for pixel-level classification task.The algorithm can obtain the original image information through the fusion of different scales of receptive field and generate the low-contrast feature map.Then the low-contrast feature map is mapped to the high-contrast feature map,and finally the high-contrast feature map reconstructs the high contrast defect image,and filter out the defects.In the screen defect detection experiments show that multi-scale model has higher classification accuracy than single-scale model.Compared with the current typical algorithms,this method has outstanding performance advantages in defect detection accuracy and speed.