| The Development of science and technology is one of the main reasons that human society gets promoted. Future society will become more technological、informational、intellectuall. With today’s information explosion, making full use of these information will contribute to the modernization of society. Image information is the most common info that people see in their daily life. Image recognition technology is an important way to use the image information. Image recognition technology also has a great demand on video surveillance and virtual reality and so on.Generally, Image target localization and recognition system includes three steps: image segmentation, target key feature extraction, target classification. After studying the background and significance of this topic, we made some research of those three steps. Deep learning theory has been widespread concerned by academia after proposed by Professor Hinton. More and more scholars try to use deep learning theory to solve the image recognition technology problems they encountered. Deep learning theory contains many models, different models are applied to different fileds. Convolutional neural networks(CNN) model is the most commonly model using for image process. Compared to traditional artificial neural network model, convolutional neural network has more hidden layers, its unique convolution and pooling operation are efficient on image processing.After analyzing the CNN model expression characteristics of image features, we build a CNN model.This model is improved based on a traditional model called VGG. The image segmentation, target key feature extraction, target classification are unified by our model. We call this model N-VGG. One of our main innovations is that after we studied the image segmentation technology of region proposal network(RPN) some suggestions were raised to improve the RPN’s efficiency. We called this model IRPN. And finally the IRPN joined into the N-VGG network. At the same time we studied the traditional activation function of convolutional neural network, and we used a new activation function which called Exponential Linear units(ELU) in our N-VGG network. Spatial pyramid pooling is also used in our N-VGG network, which can increase the recognition accuracy. And then we built this model by an open source deep learning toolbox called Caffe. A simple image target localization and recognition system is created based on this model. With this system we tested the CNN model for its accuracy. At the test section, we trained a SVM classifier, and compared it with softmax classifier to see whether it can increase the accuracy or not.Finally we made some summarization of the problems and experience during our work, and then we concluded what we can do of the image target localization recognition technology in the future. |
