This paper presents a deep learning method application to the extraction of emotions included in Chinese speech with a deep belief network (DBN) structure. Eight proper features such as pitch, mel frequency cepstrum c...
详细信息
ISBN:
(纸本)9781479974351
This paper presents a deep learning method application to the extraction of emotions included in Chinese speech with a deep belief network (DBN) structure. Eight proper features such as pitch, mel frequency cepstrum coefficient (MFCC) are chosen from Mandarin speech used as network inputs, and a DBN classifier is used instead of traditional shallow learning methods to recognition of emotions. Experiment studies have proven that its recognition rate is higher than that of the traditional back propagation (BP) method and support vector machine (SVM) classifier.
Molecular dynamics (MD) simulations are useful in various areas. In this paper, we parallelize and optimize the grid-based MD algorithm on Many Integrated Core (MIC) Architecture. To get full play of the hardware and ...
详细信息
Molecular dynamics (MD) simulations are useful in various areas. In this paper, we parallelize and optimize the grid-based MD algorithm on Many Integrated Core (MIC) Architecture. To get full play of the hardware and accelerate computation of MD simulation, we design the parallel structure using multi-threads with OpenMP. Also, various or method such as Array Notification, intrinsic and so on are used to vectorize the application according to the character of MIC for a higher performance. Due that multi-core is also a trendy of CPU and High Performance Computing, our method can be followed by other similar applications and provide a more choice.
Identifying buildings in disaster areas quickly and conveniently plays an important role in post-disaster reconstruction and disaster assessment. Aiming at the technical requirements of earthquake relief projects, thi...
详细信息
Analyzing notor imagery electrocardiogram (ECoG) signal is very challenging for it is hard to set up a classifier based on the labeled ECoG obtained in the first session and apply it to the unlabeled test data obtaine...
详细信息
Analyzing notor imagery electrocardiogram (ECoG) signal is very challenging for it is hard to set up a classifier based on the labeled ECoG obtained in the first session and apply it to the unlabeled test data obtained in the second session. Here we propose a new approach to analyze ECoG trails in the case of session-to-session transfer exists. In our approach, firstly, dimension reduction is performed with independent component analysis (ICA) decomposition. Secondly, ECoG trials are clustered by an unsupervised learning algorithm called affinity propagation. Primary experimental results show that the proposed approach gives the reasonable result than that using the classical K-means clustering algorithm.
There are two key problems in efficient large scale texture mapping for terrain rendering-efficient data organization and real time data updating in memory. In order to solve these problems, in this paper we propose a...
详细信息
There are two key problems in efficient large scale texture mapping for terrain rendering-efficient data organization and real time data updating in memory. In order to solve these problems, in this paper we propose a quadtree based indexing method to organize multi-resolution images and to fast retrieve data from disk; For memory updating, we present a real time dual-cache structure based updating method, which effectively reduces the frequency of data refresh. We also innovatively use a wavelet image enhancement algorithm to enhance original terrain texture, which obtain richer edge information and give us a more realistic effect in terrain rendering. Through the analysis of storage efficiency and rendering speed of our experiment, this dual-cache structure based method solves rendering speed and memory limit problems perfectly.
Texture classification is an important problem in image analysis. A considerable amount of research work has been done for local or global rotation invariant feature extraction for texture classification. Local invari...
详细信息
Texture classification is an important problem in image analysis. A considerable amount of research work has been done for local or global rotation invariant feature extraction for texture classification. Local invariant features contain the spatial information, but usually do not have the contrast information. A new hybrid approach is proposed which considers the contrast information in spatial domain and the phase information in frequency domain of the image. It uses the joint histogram of the two complementary features, local phase quantization (LPQ) and the contrast of the image. Support vector machine is used for classification. The experimental results on standard benchmark datasets for texture classification Brodatz and KTH-TIPS2-a show that the proposed method can achieve significant improvement compared to the LPQ, Gabor filer or local Binary pattern methods.
We propose a novel method, the complete two-dimensional principal component analysis (complete 2DPCA), for image features extraction. Compared to the original 2DPCA, complete 2DPCA not only gain a higher recognition r...
详细信息
We propose a novel method, the complete two-dimensional principal component analysis (complete 2DPCA), for image features extraction. Compared to the original 2DPCA, complete 2DPCA not only gain a higher recognition rate, but also reduce the feature coefficients needed for face recognition. Complete 2DPCA is based on 2D image matrices. Two image covariance matrices are constructed directly using the original image matrix and theirs eigenvectors are derived for image feature extraction. Our experiments were performed on ORL face database, and experimental results show that the proposed method has an encouraging performance
We propose a fast algorithm which is based on the beam let decomposition for real-time rendering of scenes in participating media with multiple scattering. Firstly, the light source radiation is considered as composed...
详细信息
We propose a fast algorithm which is based on the beam let decomposition for real-time rendering of scenes in participating media with multiple scattering. Firstly, the light source radiation is considered as composed by all particles in the media and each particle radiation is decomposed along different forward directions using the plane decomposition method. Then the multiple scattering radiation of one particle is calculated by the decomposition radiations from its adjacent particles and the light source. Finally, according to the multiple scattering radiation value of each particle, the radiation of the ray which is from viewpoint is calculated using ray marching method, which can be implemented on the graphics processing unit (GPU), and rendering process is highly parallel. The experimental results show that the algorithm can achieve real-time rendering efficiency and enhance the practicality of multiple scattering.
Recently, spatial principal component analysis of census transform histograms (PACT) was proposed to recognize instance and categories of places or scenes in an image. When combining PACT with Local difference Magnitu...
详细信息
Recently, spatial principal component analysis of census transform histograms (PACT) was proposed to recognize instance and categories of places or scenes in an image. When combining PACT with Local difference Magnitude Binary pattern (LMBP), a new representation called Local Difference Binary pattern (LDBP) was proposed and performed better. LDBP is based on the comparisons between center pixel and its neighboring pixels. However, the relationship among neighbor pixels is not considered. In this paper we proposed Local Neighbor Binary pattern (LNBP) to utilize the relationship among neighboring pixels. LNBP provides complementary information regarding neighboring pixels for LDBP. We propose to combine LDBP with LNBP, and used a spatial representation for scene recognition. Experiments on two widely used dataset demonstrate the proposed method can improve the performance of recognition.
Dimension reduction methods are often used to analyzing high dimensional data, linear dimension methods are commonly used due to their simple geometric interpretations and for effective computational cost. Dimension r...
详细信息
Dimension reduction methods are often used to analyzing high dimensional data, linear dimension methods are commonly used due to their simple geometric interpretations and for effective computational cost. Dimension reduction plays an important role for feature selection. In this paper, we have given a detailed comparison of state-of-the-art linear dimension reduction methods like principal component analysis (PCA), random projections (RP), and locality preserving projections (LPP). We have determined which dimension reduction method performs better under the FastTag image annotation framework. Experiments are conducted on three standard bench mark image datasets such as CorelSk, IAPRTC-12 and ESP game to compare the efficiency, effectiveness and also memory usage. A detailed comparison among the aforementioned dimension reduction method is given.
暂无评论