Architectural Style Classification Algorithms Research Based On Ensemble Projection And Convolution Neural Network

With the rapid development of several multimedia technologies, database technology, massive storage technology and network technology, digital images are growing at an increasing number hundreds of millions. In this huge amounts of digitial images, the research that how to organize and classify them effectively, becomes the most important work.Secondary classification based on Ensemble Projection(SEP) algorithm is proposed in this thesis to classify architectural images. At the same time, since some achievements were obtained by convolution neural network in image classification, so this thesis also uses convolution neural network to classify architecture images. The architectural style of the image classification is mainly studied in this thesis. Due to the certain similarity between different architectural style and the differences of the same architectural style with the variation of regions, there are some issues that cannot be solved by traditional algorithms in the architectural style of classification. In this thesis, these issues are solved by the two algorithms.The SEP algorithm classifies architectural style images. Because of that trees, sky and other non-building elements exist in the architectural style images, the accuracy of the image classification is reduced. In order to extract building elements of architectural style images and reduce the influence of non-building elements to the result of classification, Deformable Part-based Models(DPM) is introduced to pretreat architectural style data sets and then the SEP algorithm is used to classify the pretreatment of image in this thesis. The SEP algorithm is divided into two parts: the unsupervised feature learning process and the supervised calssification process. The unsupervised feature learning process determines sample points using Max-Min Sampling method, which uses the maximum distance between sample points to determine the different types, and uses a minimum distance between sample points to determine the same type, so the issue of the difference between the different architectural styles and the same kind of similarity between architectural styles are solved. The supervised classification process classfies images using the idea of secondary clssification, in which weighted average method is used to determine the cut-off point.In addition to the SEP algorithm, convolution neural network is adopted to classify architectural style images in this thesis. Before convolution neural network classifies images, DPM algorithm is used to preprocess image. Pretreatment of image is regarded as the input of convolution neural network that extracts composition characteristics of the underlying form characteristics of high-level abstraction, by which architectural style images were classified.Ten classes and twenty-five classes are classified by SEP algorithm and convolution neural network. Classification accuracy of ten classes and twenty-five classes architectural style images classified by SEP algorithm which combines with SVM are 0.7821 and 0.5535. Classification accuracy of SEP algorithm with LR are 0.7716 and 0.5352. Classification accuracy of ten classes and twenty-five classes architectural style images classified by Convolution neural network are 0.7380 and 0.5550.