Design And Implementation Of Deep Learning Algorithm Based On Convolutional Neural Network

With the development of So C processing performance,the progress of machine learning algorithm and the explosive increase of data,deep learning has drawn more and more attention.Convolutional Neural Network(CNN)has made great breakthrough and progress in the field of computer vision.However,most of the performance improvement depend on the increase of the convolutional layer and the expansion of the training data set,which could make complexity rise sharply.Therefore,it's important to design the CNN model with low complexity.To address this issue,a CNN model with five convolutional layers is designed to be applied to face recognition problem.In order to improve the accuracy,we add more the convolutional layers and full connection layers,increase the feature maps and full connection dimensions.To reduce the complexity,max feature map activation function is utilized to replace the rectified linear unit function.Each time the feature maps go through the activation layer,the parameters could be reduced by half in the activation layer.Moreover,network in network is introduced in the CNN model.In this network,the convolution operation is performed by a filter of size 1×1 so that the parameters of the convolution layer could be reduced by 8/9.Meanwhile,it will also enhance the generalization ability without change of the number of layers.In addition,the dimension of full connection lays is decreased to further reduce the parameters as far as possible to ensure the accuracy.Test results demonstrate that the accuracy of deep CNN model increases by 15% compared to CNN model by adding convolutional layers simply.However,the size of deep CNN model is 28 times bigger than the CNN one.After the complexity optimization of the deep CNN model,the accuracy of light CNN model increases by 5% compared to CNN model and the size of light CNN model is 1/20 the size of deep CNN model.In the meantime,light CNN model outperforms the Goog Le Net and Res Net models in terms of complexity.The test results verify the effectiveness of light CNN model in reducing the model complexity.