Plant Image Set Classification Using Reverse Trainning And Deep Learning

In image set classification, large number of images take the place of single or little images in per image set of gallery sets and probe sets, and the images of one set are belong to the same class. The main work of the image set classification algorithms is used to represent each image set with a single entity, such as exemplar image, mean, subspace, manifold, and so on. The other work is to define the similarity measure between the entities. The plant images classification is one of the very important research directions in pattern recognition and computer vision, and it's emphasized to study the plant image set classification methods. The work of this paper is focusing on studying the methods of plant images preprocessing, image feature extraction, and the plant image set classification, and the most important work is presented reverse train and deep learning for plant image set classification.The main works of this paper are as follows:(1) Verify the effective of feature fusion in plant image set classification. The accuracy rate is related to feature extraction in image set classification, and feature fusion can afford the more effective feature information, and thus the feature fusion can increase the accuracy rate. In plant image set classification, there are shape feature and texture feature of plant images that can be used to realize multi-feature fusion to improve classification accuracy.(2) Propose a plant image set classification method based on reverse training. In the view of probe set in the plant image set classification, the plantae on the earth can be divided into the two parts: the one is the same class with the probe class and the other is difference, so if we can find the hyper plane between the probe set and the other sets, we solve the problem of image set classification. The one work of this paper is to use reverse train in the plant image set classification. In this work, gallery image set are divided into the mixed-class set and the remain with the help of clustering algorithm, and then learn a classifier using SVM with a kernel function using the mixed-class set and the query set.(3) Develop a plant image set classification algorithm using a deep model. To take advantage of the learning ability of deep learning, design a specific deep learning model for the per gallery set and then initialize the parameters of per deep model using Gaussian Restricted Boltzmann Machines. Finally, the each image in the probe set will be put in the deep models, and the output is the cluster of reconstruction error from each model, and then get the information of the probe set. The accuracy and sensitive of image set classification could be improved using the deep learning.