Search

Convolutional neural networks (CNN) have achieved outstanding performance in image recognition tasks, but many hand-crafted features still play important roles in some areas. Hand-crafted features are designed to describe image content from specific aspects, which may provide complementary information for CNN in image classification tasks. This paper explores feature fusion methods and proposes a novel framework for combining CNN with hand-crafted features. The framework has two main advantages. First, feature encoder can encode non-normalized features in CNN, which takes advantage of some good edge, texture and local features. Second, joint training strategy makes features fuse better in CNN. We validate that many handcrafted features help to improve the performance of origin CNN. Experiments show our method outperforms the origin CaffeNet on Cifar10 dataset with 79.16% accuracy.

This paper describes a rose analysis and recognition system and presents the main principles used to realize the recognition system. The major principles presented in this paper are the mathematical description methods for rose features such as shape, size and color of the flower, petal, leaf, etc, and the object-oriented pattern recognition (OOPR) approach which mathematically deals with how to comprehensively use all different rose features rationally in the recognition scheme. The recognition system is described and some of its experimental results are given which demonstrate the efficiency of our methods.

This paper presents a novel pattern recognition application, the application in agronomy and proposes a roundness measurement concept to analyze the roundness of rose flower shapes for the new application. The rationality of this concept is theoretically analyzed and its mathematical definition is given. Experiment results show that this measurement is an efficient parameter for image shape analysis.

The asymptotic stability of a continuous neural network is analyzed for associative memory. An optimal design method is proposed which ensures the highest associative memory speed and guarantees the storage of each desired memory with attractivity. The network asymptotic stability is analyzed by means of a new energy function, and four theorems are obtained. By comparing these theorems with existing ones, it can be shown that in some cases they are consistent, while in others they are not equivalent but complementary to each other. Further study results in two more generalized conclusions, of which the existing conclusions are special cases. The network optimal design method is proposed in terms of an optimal associative memory theorem. Two application examples are presented to demonstrate the defeffectiveness of the optimal design method, which can be used to design the network for many applications.

Image composition is a attractive domain and there are some classical composition technologies. Possion composition and gradient domain are some typical and traditional composition technologies. Their main function is to create plausible composited image. We can use the traditional methods to get a realistic visual results. However, after the composition we found that the details is not preserved, especially for the boundary of the composited image. We use some innovative methods to deal with the problem and there are two main stages in the paper. The first stage is that we introduce the model of our filter and compare it to the other filter. The second stage is that we use the filter and image pyramid to preserve the composited image's details after the image composition, especially for the boundary.

In this paper, the structure and design criteria of a multimedia information processing and analyzing system (MIPAS), which can be used to deal with more complicated intelligent issues, are presented. According to the structure and design criteria, a software environment (SE-MIPAS) which supports the implementation of multimedia information processing and analyzing applications, is implemented and introduced. Under this software environment, a demonstration MIPAS system is constructed. Experiments show that the multimedia information processing and analyzing is much more powerful and effective than single medium case.

A new 3D model retrieval approach is proposed by using multi-instance to understand 3D models, while each object in a 3D model is called instance. For visually connected components are not geometrically connected, we present a two-stage multi-instance generation algorithm. The first stage is to compute connected components from a mesh model. The second stage is to merge connected components by testing intersections among their axis-aligned bounding boxes. And then shape descriptor is extracted for each instance and the similarity between 3D models is measured by means of instances' shape descriptors. Experimental results show that our method has better performance than the traditional 3D model retrieval approach.

It is shown that images can be subject to lossy compression in such a way that introduced distortions are not visible. For this purpose, two modern visual quality metrics, MSSIM and PSNR-HVS-M, can be used. Their values are to be provided not less than 0.99 and 40 dB, respectively, and the corresponding lossy compression is to be carried out. Attained compression ratio (CR) depends upon image properties and a coder used. The proposed methodology of lossy compression can be successfully exploited in remote sensing and medical imaging with producing CR by several times larger than the best lossless image compression techniques.

Human motion capture typically requires synchronization hardware and specialized cameras, which makes the system expensive and technically complex. Instead, in this paper, we propose a novel approach for markerless motion capture with multiple unsynchronized cameras. Cameras synchronization is achieved using sharp illumination changes produced while recording the videos. To estimate poses with the presence of sharp illumination changes, we also propose a pose estimation method based on a bilayer Gaussian Mixture Model (GMM) and graph-cut, which can handle sharp illumination changes. And we demonstrate our video synchronization and markerless motion capture methods by the experiments on real human motion videos.

Publication in the conference proceedings of EUSIPCO, Glasgow, Scotland, 2009

In this paper, we present an approach for markerless model-based full human-body motion capture using multi-view images as input. We extract volume data (voxels) representation from the silhouettes extracted from multiple-view video images by the method of shape from Silhouettes (SFS), and match our predefined human body model to the volume data. We construct an energy field in the volume of interest based on the volume data and human body model with pose parameters, and transform the matching to an energy minimizing problem. By dynamic graph cut, we get the minimum energy of certain pose parameters, and at last we optimize the pose parameters using Powell algorithm with a novel approach that uses the linear prediction guiding the optimization process and get the pose recovered. Through the test results on several video sequences of human body movements in an unaugmented office environment, we demonstrate the effectiveness and robustness of our approach.

Intelligent virtual environment (IVE) combines virtual reality and artificial intelligence (A I), i.e. It incorporates AI technology into virtual environment, which makes the virtual environment more interactive and credible. Using IVE concept, we design and implement an entertainment IVE application platform with intelligent agents and neural networks. The humanlike intelligent objects in the virtual environment are implemented with intelligent agents, which are endowed with intelligence for self-learning and collaborating with each other by applying BP and ESP neural network to the agent controllers.

This paper studies the feasibility of information analysis processing technology, which fuses speech and image together in the real-time monitoring system. It emphasizes particularly on speech analysis and fuses these two technologies in terms of scoring strategy. It also makes some improvement on MFCC feature extraction and proposes a quick MFCC algorithm. The proposed algorithm can reach the requirement of real-time system in case of the high precision. To prove it, this paper compares its algorithm with LPC and FFT. The experiment indicates that the EER of LPC is 13.9% and the EER of FFT is 11.1%, but by using the Quick MFCC the EER is only 4.2%. And compared with the traditional MFCC algorithm, the quick MFCC algorithm reduces the run time greatly while maintaining recognition accuracy of the system. Finally the rate of fusion recognition is about 97.8%, which is a good result for the real-time monitoring system.

Proposes an extended bidirectional associative memory (BAM) neural network model which can do auto- and hetero-associative memory. The theoretical proof for this neural network model's stability is given. Experiments show that this neural network model is much more powerful than the M-P model, discrete Hopfield neural network, continuous Hopfield neural network, discrete bidirectional associative memory neural network, continuous and adaptive bidirectional associative memory neural network, backpropagation neural network and optimal designed nonlinear continuous neural network. Experimental results also show that, when it does auto-associative memory, the power of this model is the same as the loop neural network model which can only do auto-associative memory.

This paper introduces the first-generation seamless messaging system of the National Research Council and describes its further development version, the multimedia seamless messaging system.

This paper describes a rose variety recognition project and proposes the modified Fourier descriptor. Based on the modified descriptor, a new measurement, angle measurement, is proposed for image shape analysis. We use it for rose flower shape analysis and the experimental results show that this measurement is quite efficient for image shape analysis.

Proposes a new associative memory neural net (NN) model called the loop neural network model, and the theoretical proof of this NN's stability is given. Experiments show that this NN model is much more powerful than the McCulloch-Pitts model, the discrete Hopfield NN, the continuous Hopfield NN, the discrete bidirectional associative memory NN, the continuous and adaptive bidirectional associative memory NN, the backpropagation NN, and the optimally designed nonlinear continuous NN. >

The advantages of using multifrequency remote sensing for measuring underlying terrain parameters and radar data interpretation are discussed. Some important peculiarities of image processing and analysis are presented and novel algorithms of two-dimensional image filtering and transformation are proposed. Examples of their application to real data are given. Possible ways of using neural networks on different stages of radar data processing and analysis are shown. >

The paper presents the structure and design principle of a neural network (NN) image understanding system which is used to recognize and analyze maps and animals with the features of translation-, scale-, and rotation-invariance. The utilized network is nonlinear continuous neural network using its associative memory function. We designed this neural network system using an optimal design method. The feature parameters which are inputted into the system to carry out the recognition task are Zernike moments. Through extensive experimentation with translation, scale, rotation as well as distortion (such as cutting off some parts of the inputted image), we can see this system is a high robustness and fault-tolerance image understanding system. >

It is shown that the array pattern parameters under the influence of spatially correlated errors and distortions can and should be optimized in a statistical sense. The proposed sum-type apodization functions (step-like functions also belong to this class) are able to provide the available trade-off of disturbed pattern characteristics. In particular, the mean maximal side lobe level (MSLL) can be reduced by 1.2 dB even in comparison with the optimal Dolph-Chebyshev and Kaiser-Bessel patterns if the initial MSLL is equal. Heuristic recommendations concerning the apodization function selection for a priori known error correlation characteristics and variance are presented.

Presents an extended loop neural network approach to handwritten character recognition. Experiments show that this method is very effective. The recognition rate by this method is higher than that by a backpropagation network. >

Presents an extended loop neural network approach to speech recognition. This speech recognition approach is characterized by the following important properties due to the associative memory neural network. (1) It has the features of great adaptivity and fault tolerance to carry out recognition. (2) The recognition system can be constructed which allows for the formation of arbitrary nonlinear decision surfaces. (3) The recognition system can perform not only the recognition task but also restore the correct information from incomplete even some extent incorrect information at the same time. Experiments are also conducted and the results show that this speech recognition approach has great application potentials. >

In this paper, the structure and design criteria of a multimedia information processing and analyzing system (MIPAS), which can be used to deal with more complicated intelligent issues, are presented. According to the structure and design criteria, a software environment (SE–MIPAS) which supports the implementation of multimedia information processing and analyzing applications, is implemented and introduced. Under this software environment, a demonstration MIPAS system is constructed. Experiments show that the multimedia information processing and analyzing is much more powerful and effective than single medium case.

In this paper, we present an extended bidirectional associative memory (BAM) neural network approach to image recognition. This approach is characterized by the following important properties due to its associative memory: 1) it has the features of great adaptivity, robustness and fault tolerance to carry out recognition; 2) the recognition system constructed allows for the formation of arbitrary nonlinear decision surfaces; and 3) the recognition system can perform not only the recognition task but also restore the correct information from incomplete even some extent incorrect information at the same time. Experiments are also conducted and the results show that this approach is very efficient and has great application potentials. >