Efficient Convolutional Neural Networks with Log-linear Kernel Density Estimation for Semantic Segmentation – We propose to utilize deep convolutional neural network (CNN) as a method for large scale image registration. Convolutional neural network is trained by training the network in a single feedforward layer, and when it is trained by convolutional layers with a single feedforward layer, it can learn the embedding vector of features. When applied to image registration, CNN trained with CNN-RNN model can achieve the best registration performance and outperforms other CNN based CNN.
Feature selection is crucial for image classification. Existing work has focused on image segmentation by using linear discriminant analysis or segmentation by multiple images. Here we propose a novel approach that uses the discriminant equation to form a segmentation problem. Specifically, the discriminant equation (DIA) is formulated as a multi-class objective function, and we show that the discriminant equation is more tractable to learn. We also show that a simple iterative approach to the formulation can be used for classifying the data in a deep architecture. Experiments show the proposed approach significantly outperforms the existing approaches.
A Tutorial on Human Activity Recognition with Deep Learning
Multilevel Approximation for Approximate Inference in Linear Complex Systems
Efficient Convolutional Neural Networks with Log-linear Kernel Density Estimation for Semantic Segmentation
Multi-point shape recognition with spatial regularization
Improving Image Classification by Leveraging the Information Co-Optimalization FrameworkFeature selection is crucial for image classification. Existing work has focused on image segmentation by using linear discriminant analysis or segmentation by multiple images. Here we propose a novel approach that uses the discriminant equation to form a segmentation problem. Specifically, the discriminant equation (DIA) is formulated as a multi-class objective function, and we show that the discriminant equation is more tractable to learn. We also show that a simple iterative approach to the formulation can be used for classifying the data in a deep architecture. Experiments show the proposed approach significantly outperforms the existing approaches.