Image Classification Based On Convolution Neural Network

Under the background of ethnic groups and culture’s diversity, national culture, as Chinese cultural treasures, are of great significance for politics, economy, culture, education. Folk cultural resources mining, as the main extraction method of national culture, is increasingly approved by academia and industry. However, traditional folk cultural resources mining methods, such as the collection and classification of folk culture festival events, need to consume a lot of human resource. Moreover, high-level precision is difficult to be maintained. Based on the questions above, image classification as the classical techniques of computer vision, could reduce the consuming of human resource and keep the precision of retrieval and classification. Compared to other huge amounts of images, the folk custom culture images are special in character and semantic. This paper mainly researches on folk culture festival event classification method of traditional folk custom culture images, set up new models on the foundation of traditional classification methods. Universality and effectiveness are verified on experimental database.The main work in this paper is as follows:(1) We proposed one convolutional neural network based, combined with target recognition and scene recognition network model of traditional folk culture festival event classification method. Since there is no specific traditional folk culture festival events database, we first set up one database based on traditional folk culture festival events. In the database,1450 learning instances and 474 testing instances are labeled by five types of labels (Lantern Festival, Arbor Day, Dragon Boat Festival, Mid-Autumn festival and National Day). For the character of traditional folk culture festival events combined with two clues of target and scene, this paper designed a convolutional neural network model of target recognition and scene recognition. This paper first adopted Fine-tuning AlexNet network training target recognition in large-scale data set Caltech256 network model, and then adopted Fine-tuning Google-Net network in the ICCV ChaLearn LAP challenge 2015 data sets on training scene recognition network model. At last, on the target training database, two pre-training results were fused by multi-model fusion method. On the target testing database, there are obvious improvements for precision after fusion, the mAP(average precision of recognition) is 84.81%. Compared to the precision before fusion, it is increased by 6.31% and 4.64% respectively.(2)We proposed a method based on the convolutional neural network, and combined with Google-Net deep convolution neural network identification model of network training target recognition network model and scene.The experiments show that the combination of soft-max classifier by traditional folk culture festival event recognition could achieve higher recognition precision obtained from the output. Moreover, the deeper the convolutional neural network structure of Google-Net in the performance of target recognition is better than that of the deep structure AlexNet by 5.28%. At the same time, using 22 convolution layer under the condition of network model can achieve the mAP of 89.87%.