PROJECT TITLE :

Semi-Supervised Deep Learning Using Pseudo Labels for Hyperspectral Image Classification - 2018

ABSTRACT:

Deep learning has gained popularity in an exceedingly variety of computer vision tasks. Recently, it's also been successfully applied for hyperspectral Image Classification tasks. Coaching Deep Neural Networks, such as a Convolutional Neural Network for classification requires a large variety of labeled samples. However, in remote sensing applications, we sometimes only have a small amount of labeled knowledge for training because they are expensive to collect, though we tend to still have abundant unlabeled data. During this Project, we tend to propose semi-supervised deep learning for hyperspectral image classification-our approach uses restricted labeled information and abundant unlabeled information to coach a deep neural network. More specifically, we tend to use deep convolutional Recurrent Neural Networks (CRNN) for hyperspectral image classification by treating each hyperspectral pixel as a spectral sequence. Within the proposed Semi-Supervised Learning framework, the abundant unlabeled data are used with their pseudo labels (cluster labels). We tend to propose to use all the coaching knowledge together with their pseudo labels to pre-train a deep CRNN, and then fine-tune using the restricted offered labeled information. Any, to utilize spatial info within the hyperspectral pictures, we propose a constrained Dirichlet method mixture model (C-DPMM), a non-parametric Bayesian Clustering algorithm, for semi-supervised clustering that includes pairwise must-link and can't-link constraints-this produces high-quality pseudo-labels, resulting in improved initialization of the deep neural network. We additionally derived a variational inference model for the C-DPMM for economical inference. Experimental results with real hyperspectral image information sets demonstrate that the proposed semi-supervised method outperforms state-of-the-art supervised and semi-supervised learning ways for hyperspectral classification.