Network Structure Optimization Of Convolutional Neural Networks In Image Classification

In recent years,the ILSVRC competition has refreshed the accuracy of image recognition year by year.The competition gave a training data set containing a large number of artificial labels,and researchers will to design their own algorithms and verify the effectiveness of their algorithms in object detection and recognition on the given data set.Another major goal is to test the advancement of computer vision technology in the retrieval and labeling of large-scale images.Many well-known network structures have emerged in the competition,such as Alex Net,VGG,Google Net,Deep Residual Network and so on.In order to make the model better applied to a specific data set,to improve the image classification accuracy as much as possible,and to reduce the oscillation and over-fitting phenomenon in the training process,this paper studied the Alex Net model in the ILSVRC competition.Main tasks as follows:1?A new threshold activation function is proposed to solve the problem that the data appearing in the training results deviate from the optimal point.The network model is more stable and the image classification accuracy is improved.The Alex Net model of the improved activation function was verified using the caltech101 dataset and the caltech256 dataset.The data training process was performed on the Caffe platform in the Linux system.Since Caffe only supports image types in the lmdb format,it is necessary to format the images before training.The results showed that the classification accuracy of the caltech101 data set increased from 0.977 to 0.993,and the classification accuracy of the caltech256 data set increased from 0.654 to 0.923.2?For the problem of slow training of threshold activation function classification,find the time complexity of network training in related literature research,and improve the Alex Net model from three aspects: the number of feature maps,the area of convolution kernels and the number of convolution layers.Under the premise of not affecting the accuracy of classification,the optimal models suitable for the classification of caltech101 dataset and caltech256 dataset were determined.The caltech101 data set achieves the goal of reducing the number of feature maps by reducing the number of convolution kernels,and the training speed is significantly improved.The caltech256 data set achieves the purpose of accelerating the convergence speed by reducing the number of convolution layers.In this paper,the important role of deep learning and convolutional neural networks in image classification is illustrated by comparing experimental data with mapping.Based on the existing convolutional neural networks,a convolutional neural network model with optimized activation function is proposed.The model improves the classification accuracy of the dataset caltech101 and the dataset caltech256,greatly enhances the network's stability and anti-over-fitting performance,and also greatly improves the convergence speed.