An Improved Deep Convolutional Neural Network And Its Weight Initialization

With the wide application of image recognition and the spring up of the deep learning,the deep convolutional neural network has been widely applied in image recognition.The deep convolutional neural network is state-of-the-art image recognition method based on feature learning.The image recognition methods based on feature learning do not need to extract the specified features.The appropriate features for classification are found by iterative learning.The identification accuracy and generalization capacity of deep convolutional neural network has been significantly improved than the traditional image recognition method.So,in recent years,the image recognition systems based on DCNN have made remarkable achievements.The main innovative work of this thesis is as follows.Convolutional neural network cannot be adequately trained on small datasets when using deep convolutional neural network for image recognition.In view of the above problems,a weight initialization method based on visual saliency and unsupervised pre-training is proposed.Firstly,the visual saliency algorithm is used to extract the salient region from the original image and some patches is cut randomly from the saliency map.Then these patches of salient regions are put into the SAE for unsupervised pre-training.Finally,the trained parameters between the input layer and hidden layer are the weights we needed.Thus,we can get the initial values of filters that consistent with the dataset statistical characteristics.Experiments on some Caltech 101 dataset and CIFAR-10 dataset prove their validity and stability.According to the dataset characteristics of the down images,a deep convolutional neural network based on the Inception module and its improved module is proposed to recognition the down types.The experimental results show that the proposed deep convolutional neural network has good recognition accuracy for the down dataset.