Research Of Image Classification Algorithm Based On Multi-scale Convolutional Neural Network

With the rapid development of the Internet,image information is exploding.Traditional image classification methods cannot be effectively used in massive data.In recent years,the rapidly developing convolutional neural network has performed well.There are two development trends of convolutional neural networks,the one is based on degradation problem of network to seek higher network accuracy by change the structure of convolutional neural networks,other is to get better performance in production by optimizing the network structure and reducing network complexity.main tasks as follows:Shallow convolutional network has a simple structure and a small amount of network parameters.With this background,the structure of the shallow convolutional neural network is improved to obtain higher accuracy.Combining the method of spatial gold tower representation image with traditional shallow convolutional neural network,this paper proposes a convolutional neural network based on spatial pyramid multi-scale feature extraction to,which has improved the accuracy of the network on small data sets.Aiming at the problem that the current deep convolutional neural network is too deep,the number of parameters is too large,and the structure hyperparameter is difficult to set,this paper proposes a new network.The network applies the method of spatial pyramid representation to the ResNeXt network,and improves the stacking module to achieve the increase or decrease of the number of overlaps.Without affecting the accuracy of the original network,the network reduces the network storage and running time by reducing the network training parameters,and is better applied to industrial production.In this paper,multi-scale feature extraction for different depth convolutional neural networks to improve network performance by studying the structure and development of convolutional neural networks.For the networks of different depth,,public data sets with different image sizes,different numbers,and different complexity were selected for testing.The experimental results show that the improved multi-scale feature extraction shallow convolutional neural network not only improves the accuracy of image classification,but also improves the adaptability of the network in non-target databases,the improved multi-scale feature extraction deep convolutional neural network greatly reduces the network parameters when the accuracy of the experimental database changes little,it makes the network to be better used in industrial production.