Mingai Li

A tabu based neural network learning algorithm

This paper presents a novel neural network learning algorithm, the tabu-based neural network learning algorithm (TBBP). In our work, the TBBP mainly use the tabu search (TS) to improve the nonlinear function approximating ability of the neural network. By using the TS in the global search, the algorithm can escape from the local minima and obtain some superior global solutions, the weights of the neural network, to approximate the nonlinear function. Results confirm that the TBBP can greatly improve the approximating ability of the neural network for several typical nonlinear functions.

A novel feature extraction method for scene recognition based on Centered Convolutional Restricted Boltzmann Machines

Scene recognition is an important research topic in computer vision, while feature extraction is a key step of scene recognition. Although classical Restricted Boltzmann Machines (RBM) can efficiently represent complicated data, it is hard to handle large images due to its complexity in computation. In this paper, a novel feature extraction method, named Centered Convolutional Restricted Boltzmann Machines (CCRBM), is proposed for scene recognition. The proposed model improves the Convolutional Restricted Boltzmann Machines (CRBM) by introducing centered factors in its learning strategy to reduce the source of instabilities. First, the visible units of the network are redefined using centered factors. Then, the hidden units are learned with a modified energy function by utilizing a distribution function, and the visible units are reconstructed using the learned hidden units. In order to achieve better generative ability, the Centered Convolutional Deep Belief Networks (CCDBN) is trained in a greedy layer-wise way. Finally, a softmax regression is incorporated for scene recognition. Extensive experimental evaluations on the datasets of natural scenes, MIT-indoor scenes, MIT-Places 205, SUN 397, Caltech 101, CIFAR-10, and NORB show that the proposed approach performs better than its counterparts in terms of stability, generalization, and discrimination. The CCDBN model is more suitable for natural scene image recognition by virtue of convolutional property.

Scene and place recognition using a hierarchical latent topic model

Abstract Place classification and object categorization are necessary functions of vision-based robotic systems. In this paper, a novel latent topic model is proposed to learn and recognize scenes and places. First, each image in the training set is characterized by a collection of local features, known as codewords, obtained by unsupervised learning, and each codeword is represented as part of a topic. Then, the codeword distribution of detected local features from the training images is learned by performing a k-means algorithm. Next, a modified Latent Dirichlet Allocation model is employed to highlight the significant features (i.e., the codewords with higher frequency in the codebook). The Highlighted Latent Dirichlet Allocation (HLDA) improves the efficiency of learning procedure. Finally, a fast variational inference algorithm for HLDA is proposed to reduce the computational complexity in parameter estimation. Experimental results using natural scenes, indoor and outdoor datasets show that the proposed HLDA method performs better than other counterparts in terms of accuracy and robustness with the variation of illumination conditions, perspectives, and scales. The Fast HLDA is order of magnitudes faster than the HLDA without obvious loss of accuracy.

Extracting the nonlinear features of motor imagery EEG using parametric t-SNE

When performing studies on brain computer interface based rehabilitation problems, researchers frequently encounter difficulty due to the curse of dimensionality and the nonlinear nature of Motor Imagery Electroencephalography (MI-EEG). Though many approaches have been proposed recently to address the feature extraction problem and have shown surprising performance, unfortunately, most of them are non-parametric or linear dimension reduction techniques, which are limited in utility for out-of-sample extension for MI-EEG classification. To address the problem and obtain accurate MI-EEG features, a new unsupervised nonlinear dimensionality reduction technique termed parametric t-Distributed Stochastic Neighbor Embedding (P. t-SNE) is employed to extract the nonlinear features from MI-EEG. Considering that MI-EEG is a kind of non-stationary signal with remarkable time-frequency rhythmic distribution characteristics, Discrete Wavelet Transform (DWT) is used to extract the time-frequency features of MI-EEG. Furthermore, P. t-SNE is applied to selected wavelet components to get the nonlinear features. They are then combined serially to construct the feature vector. Experiments are conducted on a publicly available dataset, and the experimental results show that the nonlinear features have great visualization performance with obvious clustering distribution, and the feature extraction method indicates excellent classification performance as evaluated by a support vector machine classifier. This paper suggests a manifold based technique for further analysis and classification research of MI-EEG.

Natural scene recognition based on Convolutional Neural Networks and Deep Boltzmannn Machines

Scene recognition is a significant topic in computer vision, and Deep Boltzmann Machines (DBM) is a state-of-the-art deep learning model which has been widely applied in object and hand written digit recognition. However, when the DBM is used in scene recognition, it is difficult to handle large images due to its computational complexity. In this paper, we present a deep learning method based on Convolutional Neural Networks (CNN) and DBM for scene image recognition. First, in order to categorize large images, the CNN is utilized to preprocess images for dimensional reduction. Then, regarding the preprocessed images as the input of the visible layer, the DBM model is trained using Contrastive Divergence (CD) algorithm. Finally, after extracting features by the DBM, the softmax regression is employed to perform scene recognition tasks. Since the CNN can reduce effectively image size, the proposed method can improve the computational efficiency and becomes more suitable for large image recognition. Experimental evaluations using SIFT Flow dataset and fifteen-scene dataset demonstrate that the proposed method can obtain promising results.

Combined long short-term memory based network employing wavelet coefficients for MI-EEG recognition

Motor Imagery Electroencephalography (MI-EEG) plays an important role in brain computer interface (BCI) based rehabilitation robot, and its recognition is the key problem. The Discrete Wavelet Transform (DWT) has been applied to extract the time-frequency features of MI-EEG. However, the existing EEG classifiers, such as support vector machine (SVM), linear discriminant analysis (LDA) and BP network, did not make full use of the time sequence information in time-frequency features, the resulting recognition performance were not very ideal. In this paper, a Long Short-Term Memory (LSTM) based recurrent Neural Network (RNN) is integrated with Discrete Wavelet Transform (DWT) to yield a novel recognition method, denoted as DWT-LSTM. DWT is applied to analyze the each channel of MI-EEG and extract its effective wavelet coefficients, representing the time-frequency features. Then a LSTM based RNN is used as a classifier for the patten recognition of observed MI-EEG data. Experiments are conducted on a publicly available dataset, and the 5-fold cross validation experimental results show that DWT-LSTM yields relatively higher classification accuracies compared to the existing approaches. This is helpful for the further research and application of RNN in processing of MI-EEG.

Applying a Locally Linear Embedding Algorithm for Feature Extraction and Visualization of MI-EEG

Robotic-assisted rehabilitation system based on Brain-Computer Interface (BCI) is an applicable solution for stroke survivors with a poorly functioning hemiparetic arm. The key technique for rehabilitation system is the feature extraction of Motor Imagery Electroencephalography (MI-EEG), which is a nonlinear time-varying and nonstationary signal with remarkable time-frequency characteristic. Though a few people have made efforts to explore the nonlinear nature from the perspective of manifold learning, they hardly take into full account both time-frequency feature and nonlinear nature. In this paper, a novel feature extraction method is proposed based on the Locally Linear Embedding (LLE) algorithm and DWT. The multiscale multiresolution analysis is implemented for MI-EEG by DWT. LLE is applied to the approximation components to extract the nonlinear features, and the statistics of the detail components are calculated to obtain the time-frequency features. Then, the two features are combined serially. A backpropagation neural network is optimized by genetic algorithm and employed as a classifier to evaluate the effectiveness of the proposed method. The experiment results of 10-fold cross validation on a public BCI Competition dataset show that the nonlinear features visually display obvious clustering distribution and the fused features improve the classification accuracy and stability. This paper successfully achieves application of manifold learning in BCI.

Semantic Map Building Based on Object Detection for Indoor Navigation

Building a map of the environment is a prerequisite for mobile robot navigation. In this paper, we present a semantic map building method for indoor navigation of a robot using only the image seque...

The novel recognition method with Optimal Wavelet Packet and LSTM based Recurrent Neural Network

In order to adaptively extract the subject-based time-frequency features of motor imagery EEG (MI-EEG) and make full use of the sequential information hidden in MI-EEG features, a Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN) is integrated with Optimal Wavelet Packet Transform (OWPT) to yield a novel recognition method, denoted as OWLR. Firstly, OWPT is applied to each channel of MI-EEG, and the improved distance criterion is used to find the optimal wavelet packet subspaces, whose coefficients are further selected as the time-frequency features of MI-EEG. Finally, a LSTM based RNN is used for classifying MI-EEG features. Experiments are conducted on a publicly available dataset, and the 5-fold cross validation experimental results show that OWLR yields relatively higher classification accuracies compared to the existing approaches. This is helpful for the future research and application of RNN in processing of MI-EEG.

Salient object detection based on global multi-scale superpixel contrast

Salient object detection, as a necessary step of many computer vision applications, has attracted extensive attention in recent years. A novel salient object detection method is proposed based on multi-superpixel-scale contrast. Saliency value of each superpixel is measured with a global score, which is computed using the regions colour contrast and the spatial distances to all other regions in the image. High-level information is also incorporated to improve the performance, and the saliency maps are fused across multiple levels to yield a reliable final result using the modified multi-layer cellular automata. The proposed algorithm is evaluated and compared with five state-of-the-art approaches on three publicly standard datasets. Both quantitative and qualitative experimental results demonstrate the effectiveness and efficiency of the proposed method.