He Xiaohai

Online forum post opinion classification based on tree conditional random fields model

There is a major defect when using the traditional topic-opinion model for post opinion classifications in an online forum discussion. The accuracy of the classification based on the topic-opinion model highly depends on the observable topic-opinion features aiming at the subject, while a large number of posts do not have such features in a forum. Therefore, for the most part, the accuracy is less than 78%. To solve this problem, we propose a new method to identify post opinions based on the Tree Conditional Random Fields (T-CRFs) model. First, we select the topic-opinion features of the posts and associated opinion features between posts to construct the T-CRFs model, and then we use the T-CRFs model to label the opinions of the tree-structured posts under the same topic iteratively to reach a maximum joint probability. To reduce the training cost, we design a simplified tree diagram module and some feature templates. Experimental results suggest the proposed method costs less training time and improves the accuracy by 11%.

Triplet circular Hough transform for circle detection

A new method, triplet circular Hough transform, is proposed for circle detection in image processing and pattern recognition. In the method, a curve in an image is first detected. Next, a sequence of three points on the curve are selected, a sequence of parameters (a, b, r) corresponding to the three points are calculated by solving the circle equation of the curve, and two 2-D accumulators A(a, b) and R(a, b) are accumulated with 1 and r, respectively. Then the parameters {(a, b, r)} of the circles fitting the curve are determined from A(a, b) and R(a, b) by searching for the local maximum over A(a, b). Because no computation loops over center (a, b) and/or radius r are needed, the method is faster than the basic and directional gradient methods. It needs also much smaller memory for accumulation.

Impact of user influence on information multi-step communication in a micro-blog

User influence is generally considered as one of the most critical factors that affect information cascading spreading. Based on this common assumption, this paper proposes a theoretical model to examine user influence on the information multi-step communication in a micro-blog. The multi-steps of information communication are divided into first-step and non-first-step, and user influence is classified into five dimensions. Actual data from the Sina micro-blog is collected to construct the model by means of an approach based on structural equations that uses the Partial Least Squares (PLS) technique. Our experimental results indicate that the dimensions of the number of fans and their authority significantly impact the information of first-step communication. Leader rank has a positive impact on both first-step and non-first-step communication. Moreover, global centrality and weight of friends are positively related to the information non-first-step communication, but authority is found to have much less relation to it.

Automated image analysis system for etched track counting—A preliminary study

Abstract A preliminary study on an automated image analysis system for etched track counting is described in this short report. The system made a feature of full automation in the operation mode of automatic measurement. All the hardware components were commercially available except the specially-designed microscope control unit. The programs for autofocus and autoscanning, and for track measurement were all developed by our group. Tests with chemically etched α -particle tracks in CR-39 detectors showed that the agreement between the results of visual counting and that of automatic counting was acceptable. Improvement of software for track measurement is still in progress and the preliminary results are encouraging.

Learning-based super-dimension (SD) reconstruction of porous media from a single two-dimensional image

The three-dimensional (3D) microscopic pore structure of Reservoir rock directly affects its seepage characteristics and physical properties. A 3D microscopic pore structure can be reconstructed from a single two-dimensional (2D) training image (TI) by using mathematical modeling methods. In this paper, we introduce the concepts of blocks, dictionary and learning into the reconstruction of 3D porous media from the area of example-based super-resolution (SR) reconstruction, and put forward the concept of super-dimension (SD) reconstruction: study the corresponding relations between 2D images and 3D images of real microscopic pore structure of reservoir rock, and use these relations as guidance for the reconstructions of a new 2D image. According to the concept of SD reconstruction, we put forward a new learning-based super-dimension (LBSD) reconstruction algorithm whose basic steps are as follows: (1) Select the training set; (2) build the dictionary; (3) reconstruction. Based on these steps, we did experiments on reconstruction of porous media from a single two-dimensional image. Comprehensive tests show that the reconstructed 3D structure consists with the 3D Micro-CT core sample where the 2D TI is selected from both in morphological characteristics and Statistical characteristics.

Low bit-rates image coding via wavelet domain downsampling and dual wavelet-pixel domain reconstruction

At low bit-rates, the conventional image coding standards, e.g., JPEG and JPEG 2000, do not have good compression performance due to the insufficiency of coding bits. A common solution to this problem is downsampling before encoding and reconstruction after decoding. Inspired by the wavelet domain downsampling-based compression scheme, we establish an enhanced low bit-rates coding framework by making the following improvements. Firstly, a regression priors-based coding artifacts reduction (RCAR) method is incorporated to preprocess the decoded low-resolution (LR) image; secondly, given the decoded low frequency wavelet coefficients, we propose to estimate its corresponding high frequency wavelet coefficients by using the joint optimized regressors (JOR) model to recover more information lost in downsampling phase; finally, the effective group-based sparse representation (GSR) model, which exploits both the nonlocal self-similarity and local sparsity properties, is utilized to perform soft decoding on the result of wavelet reconstruction. Experimental results suggest that the proposed framework outperforms JPEG 2000 at low to medium bit-rates in terms of both quantitative and visual comparisons.