Research On The Application Of Deep Learning In Image Clustering And Classification

In the nearest decade,the development of computer software,hardware technology,smart phones and many other electronical equipments lead to the aggregated image data,which makes it necessary to mine the potential models and rules embedded in image data.Therefore the importance of data mining and machine learning technology is increasing in image big data analyses.The successful applications of deep learning in image recognition,image retrieval and image generation make it an indispensable tool for image big data analyses.By nonlinear mapping of the original data and learning network parameters,deep learning can match complex functions,and obtain the feature representation of whole data set from a small number of samples.Deep learning can automatically adjust network parameters to obtain the feature expression of original data as close as possible.By combining the low dimensional representation of data,the abstract high-dimensional representation,attribute category or feature can be obtained,and the hierarchical representation of data can be obtained too,so as to replace manual feature extraction.The depth model of deep learning and its high matching to data make it suitable to analyze large-scale data.Considering the advantages of deep learning,we adopt it into image clustering and image localization and recognition,which not only realizes the automatic extraction of image features,but also improves the accuracy of image clustering and classification.This thesis first advance the network structure to overcome the feature loss weakness of DCEC,and propose a new feature extraction method.Then aiming at the butterfly recognition problem of butterfly photos in ecological environment,we propose to introduce the deep learning based object detection method to butterfly recognition in the data set of butterfly photos in ecological environment.Our method realize the location detection and categorical recognition of butterflies in ecological butterfly photos simultaneously.The main work are as follows:1.An unsupervised deep clustering algorithm with 17 layers of network structure is proposed,which improves the weakness of DCEC algorithm in image feature loss in the training process,and combines the advantages of preserving local structure of IDEC algorithm.Two cost loss functions of DEC and IDEC are used to optimize the model,and two optimization methods of Adam and Mini-Batch SGD are used to adjust the model parameters.For complex data sets,the Inception model is first used to extract features,then the proposed deep clustering algorithm is used to do clustering.The experimental results on three image data sets demonstrate that the proposed deep network structure is robust and universal to image features,and the proposed feature extraction method is necessary and effective.2.The deep learning based target detection algorithm Faster-Rcnn is applied to location detection and specie identification of butterflies in their ecological photographs with small number of samples.The training data set is constructed in two ways,and is expanded.The mAP index is adopted to value the effect of butterfly location detection and specie recognition by Faster-Rcnn.The experimental results disclose that the deep learning based target detection algorithm Faster-Rcnn can successfully detect the locations of butterflies in ecological photographs and recognize their species simultaneously.This fact reveals that deep learning technology can be applied to fine-grained and small-sample classification problems.