Da Pan

An approach to automatic summarization for Chinese text based on the combination of spectral clustering and LexRank

In the past half century, automatic summarization has been a hot topic in the field of natural language processing, and it will be paid more and more attention to with the rapid development of the mobile network technology. Most of the automatic summarization research today is on extractive summarization, which mainly ranks the sentences according to their simple heuristic features such as the frequency of words they contain, their position in the text or paragraph and so on. Inspired by the great performance of LexRank, we manage to introduce LexRank to Chinese texts. In order to make up the deficiency of LexRank, spectral clustering is adopted to process the component analysis. All in all, we propose an approach of extractive summarization for Chinese text based on the combination of spectral clustering and LexRank, which is of high coverage and low redundancy. It is demonstrated by experiments that our approach has been greatly improved compared to the original LexRank. In addition, our approach is easy to implement and robust to noise.

The recognition of Chinese caption text in news video using convolutional neural network

News video caption, which carries main contents of related news story, plays an important role in content-based video analysis and retrieval system. In this paper, the convolutional neural network (CNN) is used to the recognition of chinese caption text in news video. First, the color and edge feature are used for caption location. Then, the segmentation combined Otsu and K-means clustering algorithm is applied to the caption images before they are sent to CNN. It is worth mentioning that we present a method for generating and labeling training images automatically, which avoids the complex and time consuming data collection. Finally, two CNN models trained on different dataset are evaluated in our experiment. By using the baseline model, the recognition accuracy can achieve 93.3% in top-1 and 98.58% in top-5 on chinese caption texts collected from news video. We also show an improvement to 95% in top-1 accuracy by averaging the two CNN models. Experimental results suggest that CNN is competent to the challenging task of chinese character recognition.

Sketch-based image retrieval by using saliency

A method of sketch-based image retrieval by using saliency is proposed. This paper introduces saliency based on the color contrast into the sketch-based image retrieval to solve the common problems of scale and translation. Moreover, an improved Hausdorff distance is proposed to increase punishment on outliers, which avoids the negative effects caused by the outliers. The experiment result shows that the proposed method has better performance in retrieval accuracy than HD-MHD and TLS-HD.

Contrast Changed Image Quality Assessment Based on Dual-Path Feature-Difference Network

Contrast plays an important role in human visual perception while it is usually unsatisfactory due to various factors in the acquisition process of images. With numerous approaches proposed to enhance contrast, less work has been dedicated to contrast changed image quality assessment (IQA). In this work, we propose a quality assessment model based on dual-path feature-difference, which uses a dual-channel convolution neural network (CNN) to extract features, removing the common feature of reference image and distorted image and preserving the feature-difference caused by contrast change. Then, image quality scores are estimated based on feature-difference maps. Moreover, a patch contrast quality map can also be created during the score prediction process, which is useful for the optimization of the image contrast enhancement algorithms. Validations based on multiple publicly available datasets show that the proposed method is well correlated with subjective evaluations of contrast changed image quality

Human pose estimation combined with depth information

Human pose estimation is an important research topic in the field of computer vision. Pose recognition is widely used in human-computer interaction, games, security, telepresence, medical and other fields. There has been some scholars put forward many new methods on this issue. But due to the quality differences between imaging devices, complicated appearance of characters, changeable body posture, a lot of uncertain factors such as interference of environmental background, human pose estimation is always one of the difficult problem in this area. In this paper, according to the existing problem of the traditional pose estimation method based on color image: manual intervention will bring the uncertainty leading to inaccurate foreground / background segmentation, and have a negative impact on the subsequent parsing process. We put forward a method which combined with the depth information captured by the Kinect, with the pre-processing of depth image, to improve pose estimation process in static images, so that the subsequent processing can be more convenient, and meanwhile the accuracy can be improved.

Coin detection and recognition in the natural scene

Although there exists some methods for coin detection and recognition, it is still a challenging task, especially for coins in natural scene. This paper proposed a method to detect and recognize the coins in natural scene. In the detection part, the Hough detection method is applied to detect the coin areas in the images. Then radius ratio, color feature and relative position constraints are used to eliminate the noise circles. In the recognition part, a multilayer convolutional neural network is used to classify proposals and get the final recognition result. Experimental results show that the proposed method could successfully detect and recognize coins in the given images.

A hybrid-architecture retrieval system based on Web Services

In this paper, we design and implement a high quality audio material retrieval system in view of the shortcomings of existing professional audio retrieval platforms. The aim of our system is to provide an integrated audio retrieval solution for audio editing staffs. Firstly, for hybrid architecture, the server of the system described in this paper publishes services by Web Services when realizing B/S logical transaction. And the rich client of C/S calls the advertised services to provide Internet audio retrieval and other functions. Then based on this system architecture, we design the retrieval and the other peripheral modules to achieve overall functions of the system. Finally, some performance tests are done to our system.