The Study Of Convolution Neural Network Based On Caffe Deep Learning Framework

Deep learning, as a significant part of artificial intelligence, has made an array of breakthroughs in many fields, such as image recognition, speech recognition, natural language processing,etc.Although which time of deep learning research is not long. In addition, it also made some gratifying achievements on applied aspects, which was not easy to deal with for traditional algorithms, including automatic unmanned vehicles, automatic pattern recognition, automatic simultaneous interpreting, commodities’ s image retrieval, handwritten character recognition and automatic license plate recognition etc.. In the past years, with the growing requirement of the development progress of deep learning for researchers, traditional deep learning programming method already could not meet the current needs, which cost researchers few months or even several years to realize the basic algorithm. For this reason, a variety of deep learning development frameworks, such as the Caffe deep learning framework in this paper, arise from the circumstance that a number of the world’s top research institutions have started seeking quick and efficient deep learning development modes. These deep learning frameworks not only provide efficient and convenient development modes for the scientific research institutions and associated developers, but some of them also offer multiple convolutional neural network models so as the study on more advanced and improved convolutional neural network model could be made by developers.Several tasks and researches are conducted based on convolution neural network of Caffe deep learning framework in the paper.First, this paper introduces the research status of deep learning in image recognition, speech recognition and natural language processing, and have a comparison within a few of mainstream deep learning frameworks, leading to Caffe deep learning framework. Artificial neural network(ANN) leads to convolutional neural network(CNN).Second, the convolution neural network elements and structure were described in detail after a discussion on artificial neutral network. Several characteristics of the Caffe framework are demonstrated and the steps of how to build Caffe environment are explained.In the end, the paper uses the Caffe deep learning framework to carry out several simulation experiments which involve following three parts: first, taking the CIFAR-10 neural network as an example, it shows how to configure the given training convolution neural network in Caffe framework; second, in accordance with the small data set, it introduces the training method of creating their own data sets and convolutional neural network; third, this paper improves the LeNet-5 network based on MNIST handwriting-characters set under the Caffe framework and activation function replaceing the original Sigmoid with ReLU function and inserting a layer of activation function into LeNet-5. The convergence speed under the proposed method to a certain extent is improved through comparison, so does the accuracy of the network training.