Optimization Of Convolutional Neural Networks Based On Unsupervised Learning And Multi-sampling

Image classification is very important for the development of research fields such as artificial intelligence and computer vision.In recent years,convolutional neural networks occupy a dominant posi tion in image classification research.However,the convolutional neural network still faces several major problems,mainly in two side: first,the features extracted by the convolution operation are easily lost during the pooling operation when passing through the network;second,the initialization value of the network parameters and design of the network structure can have a large impact on network performance.In this paper,the convolutional neural network development system is studied and optimized around the problem that the image classification algorithm is not accurate and the speed is not fast.The main contributions include:1.Optimization of multi-sampling pooling for convolutional neural networks.Aiming at the problem that the traditional pooling method easily ignores the effective information in the feature map,this paper studies how to reduce the training parameters while retaining the detailed features in the sampling process,and proposes the optimization scheme of the multi-sampling pooling for the convolutional neural network.The scheme can extract the comprehensive and effective image features by constructing the basis function to achieve the effect of improving the network recognition rate.2.The automatic encoder optimizes the convolutional neural network parameters.Aiming at the problem that the initial value state of the convolutional neural network parameters has a great impact on the network performance,based on the previous research,the stack convolutional automatic encoder is used to initialize the network parameters.The data set is pre-trained by the stack convolutional automatic encoder to extract useful features.Since the extracted feature is an implicit representation of the stack convolutional automatic encoder,the feature is used as the initialization value of the convolutional neural network parameter enables the network to train in a better direction.3.Multi-scale depth separable convolution learning network.Inspired by MobileNet,we propose a multi-scale learning network and apply it to image classification for the problem of large convolutional neural network parameters and large computational complexity.This method has two advantages:(1)it uses multi-scale blocks with deeply separable convolutions,which allows multiple sub-networks to be formed within the same model;(2)it combines multi-scale blocks with residual structures to significantly accelerate the training of the network.The three optimization methods from different angles,looking for different solutions to the corresponding problems,to improve network performance as the fundamental starting point.The proposed convolutional neural network optimization method was tested on the open dataset in the field of image classification.The experimental results show that the optimization of convolutional neural network by multi-sampling pooling speeds up the training speed of convolutional neural network and improves its testing accuracy.The method of initializing the convolutional neural network by convolutional automatic encoder can enhance the stability the network model,its test accuracy is better than the commonly used Xavier algorithm and MSRA algorithm;multi-scale depth separable convolution learning network can find a suitable balance between network performance and training volume,so that convolutional neural network can adapt to more application scenario.