Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302190
C. Chu, Zi-Xiang Kuai, Yuemin M. Zhu
In diffusion magnetic resonance imaging, the fiber tractography generally desires the estimation of intravoxel multiple fiber orientations (MFOs) with high accuracy and reliability. In general, spherical deconvolution (SD) based methods have many advantages for MFOs estimation. However, these methods are lowly immune to noise. To cope with this problem, regularization techniques were introduced in SD-based methods to reduce noise artifacts. But, the regularizers were often defined as a convex function to make the model resolving simpler, which limits their effect of regularization. In this work, we introduce a nonconvex regularizer in the Richardson-Lucy based SD framework for estimating MFOs. The results on synthetic phantom and physical phantom images demonstrate that the proposed method is superior to existing SD-based methods in terms of mean angular errors, edge preservation and computation time.
{"title":"Estimation of multiple fiber orientations using nonconvex regularized spherical deconvolution","authors":"C. Chu, Zi-Xiang Kuai, Yuemin M. Zhu","doi":"10.1109/CISP-BMEI.2017.8302190","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302190","url":null,"abstract":"In diffusion magnetic resonance imaging, the fiber tractography generally desires the estimation of intravoxel multiple fiber orientations (MFOs) with high accuracy and reliability. In general, spherical deconvolution (SD) based methods have many advantages for MFOs estimation. However, these methods are lowly immune to noise. To cope with this problem, regularization techniques were introduced in SD-based methods to reduce noise artifacts. But, the regularizers were often defined as a convex function to make the model resolving simpler, which limits their effect of regularization. In this work, we introduce a nonconvex regularizer in the Richardson-Lucy based SD framework for estimating MFOs. The results on synthetic phantom and physical phantom images demonstrate that the proposed method is superior to existing SD-based methods in terms of mean angular errors, edge preservation and computation time.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"26 1","pages":"1-5"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"89296377","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8301988
Chaoyi Chen, Xiaoqi Li, Bin Zhang
The development of the Internet has led to the accumulation of a large number of images in various databases. People are eager to find useful information in these databases which stimulate the development of image retrieval technologies. In this paper, we mainly study image retrieval based on the convolutional neural network. The study is divided into four parts to explore characteristics of convolution neural networks used in image retrieval. The first part introduces the structure of the convolutional neural network and the method of extracting features from images. The second part compares the effects of different similarity measures on retrieval accuracy. The third part studies the way to speed up retrieval. We use PCA to reduce feature dimensions and draw a line chart of dimension and accuracy. Then we analyze the reason why the change of accuracy rate is divided into two stages: ascending first and descending later. The fourth part studies the way to increase retrieval accuracy. We compare the retrieval accuracy before and after fine-tuning and analyze the reasons for that. In the end, we sum up the whole text and summarize key points that we should consider when designing an image retrieval system based on the convolutional neural network.
{"title":"Research on image retrieval based on the convolutional neural network","authors":"Chaoyi Chen, Xiaoqi Li, Bin Zhang","doi":"10.1109/CISP-BMEI.2017.8301988","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8301988","url":null,"abstract":"The development of the Internet has led to the accumulation of a large number of images in various databases. People are eager to find useful information in these databases which stimulate the development of image retrieval technologies. In this paper, we mainly study image retrieval based on the convolutional neural network. The study is divided into four parts to explore characteristics of convolution neural networks used in image retrieval. The first part introduces the structure of the convolutional neural network and the method of extracting features from images. The second part compares the effects of different similarity measures on retrieval accuracy. The third part studies the way to speed up retrieval. We use PCA to reduce feature dimensions and draw a line chart of dimension and accuracy. Then we analyze the reason why the change of accuracy rate is divided into two stages: ascending first and descending later. The fourth part studies the way to increase retrieval accuracy. We compare the retrieval accuracy before and after fine-tuning and analyze the reasons for that. In the end, we sum up the whole text and summarize key points that we should consider when designing an image retrieval system based on the convolutional neural network.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"19 1","pages":"1-5"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"89556541","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302032
Yiyou Guo, H. Huo, T. Fang
With the increasing application of high-resolution remote sensing image, image categorization becomes a more and more important technique. Recently, Convolution Neural Network (CNN) has been widely used in various computer vision tasks, for instance, generic image recognition, object detection and image segmentation. A key factor which influences the performance of CNN is the large quantity of the training images. However, it is hard to obtain large amounts of high-resolution quality images while domain adaptation can be adopted in solving this issue. As a result, in this work, we exploit domain adaptation-based CNN into high-resolution image classification task. Experiments are carried out on a latest large remote sensing image benchmark dataset. Extensive results prove the effectiveness of the proposed model.
{"title":"Remote sensing image categorization with domain adaptation-based convolution neural network","authors":"Yiyou Guo, H. Huo, T. Fang","doi":"10.1109/CISP-BMEI.2017.8302032","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302032","url":null,"abstract":"With the increasing application of high-resolution remote sensing image, image categorization becomes a more and more important technique. Recently, Convolution Neural Network (CNN) has been widely used in various computer vision tasks, for instance, generic image recognition, object detection and image segmentation. A key factor which influences the performance of CNN is the large quantity of the training images. However, it is hard to obtain large amounts of high-resolution quality images while domain adaptation can be adopted in solving this issue. As a result, in this work, we exploit domain adaptation-based CNN into high-resolution image classification task. Experiments are carried out on a latest large remote sensing image benchmark dataset. Extensive results prove the effectiveness of the proposed model.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"1 1","pages":"1-5"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"86754127","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302281
D. Cheng, Manhua Liu
Accurate and early diagnosis of Alzheimer's disease (AD) plays a significant part for the patient care and development of future treatment. Magnetic Resonance Image (MRI) and Positron Emission Tomography (PET) neuroimages are effective modalities that can help physicians to diagnose AD. In past few years, machine-learning algorithm have been widely studied on the analyses for multi-modality neuroimages in quantitation evaluation and computer-aided-diagnosis (CAD) of AD. Most existing methods extract the hand-craft features after image preprocessing such as registration, segmentation and feature extraction, and then train a classifier to distinguish AD from other groups. This paper proposes to construct multi-level convolutional neural networks (CNNs) to gradually learn and combine the multi-modality features for AD classification using MRI and PET images. First, the deep 3D-CNNs are constructed to transform the whole brain information into compact high-level features for each modality. Then, a 2D CNNs is cascaded to ensemble the high-level features for image classification. The proposed method can automatically learn the generic features from MRI and PET imaging data for AD classification. No rigid image registration and segmentation are performed on the brain images. Our proposed method is evaluated on the baseline MRI and PET images from Alzheimer's Disease Neuroimaging Initiative (ADNI) database on 193 subjects including 93 Alzheimer's disease (AD) subjects and 100 normal controls (NC) subjects. Experimental results and comparison show that the proposed method achieves an accuracy of 89.64% for classification of AD vs. NC, demonstrating the promising classification performance.
{"title":"CNNs based multi-modality classification for AD diagnosis","authors":"D. Cheng, Manhua Liu","doi":"10.1109/CISP-BMEI.2017.8302281","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302281","url":null,"abstract":"Accurate and early diagnosis of Alzheimer's disease (AD) plays a significant part for the patient care and development of future treatment. Magnetic Resonance Image (MRI) and Positron Emission Tomography (PET) neuroimages are effective modalities that can help physicians to diagnose AD. In past few years, machine-learning algorithm have been widely studied on the analyses for multi-modality neuroimages in quantitation evaluation and computer-aided-diagnosis (CAD) of AD. Most existing methods extract the hand-craft features after image preprocessing such as registration, segmentation and feature extraction, and then train a classifier to distinguish AD from other groups. This paper proposes to construct multi-level convolutional neural networks (CNNs) to gradually learn and combine the multi-modality features for AD classification using MRI and PET images. First, the deep 3D-CNNs are constructed to transform the whole brain information into compact high-level features for each modality. Then, a 2D CNNs is cascaded to ensemble the high-level features for image classification. The proposed method can automatically learn the generic features from MRI and PET imaging data for AD classification. No rigid image registration and segmentation are performed on the brain images. Our proposed method is evaluated on the baseline MRI and PET images from Alzheimer's Disease Neuroimaging Initiative (ADNI) database on 193 subjects including 93 Alzheimer's disease (AD) subjects and 100 normal controls (NC) subjects. Experimental results and comparison show that the proposed method achieves an accuracy of 89.64% for classification of AD vs. NC, demonstrating the promising classification performance.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"26 1","pages":"1-5"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"86773303","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302317
Yonas Gashaw, Fang Liu
In today's information era, the Internet is a powerful platform as the data repository that plays a great role in storing, sharing, and retrieve information for knowledge discovery. However, as there are countless, dynamic, and significant growth of data, web users face big problems in terms of the relevant information required. Consequently, poor information precision and retrieval are part of the hottest recent research areas in today's world. Despite the voluminous of information resided on the web, valuable informative knowledge could possibly be discovered with the application of advanced data mining techniques. Association rule mining, as a technique in data mining, is one way to discover frequent patterns from various data sources. In this paper, three of the foremost association rule mining algorithms used for frequent pattern discovering namely, Eclat, Apriori, and FP-Growth examined on three sets of transactional databases devised from server access log file. The comparison is made both in execution time and memory usage aspects. Unlike most previous research works, findings, in this paper, reveal that each of the algorithms has their own appropriateness and specificities that can best fit depending on the data size and support parameter thresholds.
{"title":"Performance evaluation of frequent pattern mining algorithms using web log data for web usage mining","authors":"Yonas Gashaw, Fang Liu","doi":"10.1109/CISP-BMEI.2017.8302317","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302317","url":null,"abstract":"In today's information era, the Internet is a powerful platform as the data repository that plays a great role in storing, sharing, and retrieve information for knowledge discovery. However, as there are countless, dynamic, and significant growth of data, web users face big problems in terms of the relevant information required. Consequently, poor information precision and retrieval are part of the hottest recent research areas in today's world. Despite the voluminous of information resided on the web, valuable informative knowledge could possibly be discovered with the application of advanced data mining techniques. Association rule mining, as a technique in data mining, is one way to discover frequent patterns from various data sources. In this paper, three of the foremost association rule mining algorithms used for frequent pattern discovering namely, Eclat, Apriori, and FP-Growth examined on three sets of transactional databases devised from server access log file. The comparison is made both in execution time and memory usage aspects. Unlike most previous research works, findings, in this paper, reveal that each of the algorithms has their own appropriateness and specificities that can best fit depending on the data size and support parameter thresholds.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"43 1","pages":"1-5"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"85884300","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302015
Z. Fang, Xinjun Chen
Statoliths are a pair of calcareous structures which can provide biological and ecological information for cephalopods. Understanding their shape will help us know the taxonomy of cephalopods, even to the species level. Ommastrephes bartramii, Dosidicus gigas and Illex argentinus are chosen to compare the shape of their statolith as a means of species identification because of their ecological importance to the marine ecosystems. The results show that D. gigas has a relatively large sized statolith and I. argentinus has the smallest. The four main parts of the statolith (dorsal dome, lateral dome, rostrum and wing) in the different species were diverse and distinguishable. The traditional method effectively separated the three species of statolith with high classification rates (92.0%–100%) by six morphology variables. The outline method produced a relatively low classification rate (73.4%–94.2%) using six harmonic numbers and stepwise discriminant analysis (SDA). The result in this study demonstrates that traditional method would achieve a better performance when the species are not so closely related phylogenetically, and outline method is more suitable for the statolith identification at the genus level. It is necessary to compare other cephalopod statoliths by different methods and find a suitable one.
{"title":"Statolith-based species identification methods for ommastrephidae species","authors":"Z. Fang, Xinjun Chen","doi":"10.1109/CISP-BMEI.2017.8302015","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302015","url":null,"abstract":"Statoliths are a pair of calcareous structures which can provide biological and ecological information for cephalopods. Understanding their shape will help us know the taxonomy of cephalopods, even to the species level. Ommastrephes bartramii, Dosidicus gigas and Illex argentinus are chosen to compare the shape of their statolith as a means of species identification because of their ecological importance to the marine ecosystems. The results show that D. gigas has a relatively large sized statolith and I. argentinus has the smallest. The four main parts of the statolith (dorsal dome, lateral dome, rostrum and wing) in the different species were diverse and distinguishable. The traditional method effectively separated the three species of statolith with high classification rates (92.0%–100%) by six morphology variables. The outline method produced a relatively low classification rate (73.4%–94.2%) using six harmonic numbers and stepwise discriminant analysis (SDA). The result in this study demonstrates that traditional method would achieve a better performance when the species are not so closely related phylogenetically, and outline method is more suitable for the statolith identification at the genus level. It is necessary to compare other cephalopod statoliths by different methods and find a suitable one.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"34 1","pages":"1-6"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"86379579","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302023
Hua Cai, Guangqiu Chen, Zhi Liu, Z. Geng
In view of the problem of the less correlative infrared(IFR) and visible(VI) images, a fusion methodology using joint bilateral filter (JBF) in the domain of non-subsampled contourlet transform (NSCT) is put forword. First, the images to be fused are divided into some sub-bands by NSCT. Then, the local window energy and the coefficient absolute value is regarded as the activity measure of approximate and detail sub-bands respectively. Decision maps are obtained by selecting max activity measure. Source images are regarded as the guided images and decision maps are used as input images in JBF. After filtering operation by JBF, the output images are treated as weight maps. The sub-band coefficients are fused by weighted average algorithm. Finally, the fused sub-bands are composed into a fused image by the inverse NSCT. The experiments on the IFR and VI image are carried out. For the assessment of fusion results, subjective and objective assessment methods are adopted. The results show that the proposed methodology can get better performance than some classical fusion method existed in the published literatures.
{"title":"Fusion algorithm of infrared and visible images based on joint bilateral filter","authors":"Hua Cai, Guangqiu Chen, Zhi Liu, Z. Geng","doi":"10.1109/CISP-BMEI.2017.8302023","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302023","url":null,"abstract":"In view of the problem of the less correlative infrared(IFR) and visible(VI) images, a fusion methodology using joint bilateral filter (JBF) in the domain of non-subsampled contourlet transform (NSCT) is put forword. First, the images to be fused are divided into some sub-bands by NSCT. Then, the local window energy and the coefficient absolute value is regarded as the activity measure of approximate and detail sub-bands respectively. Decision maps are obtained by selecting max activity measure. Source images are regarded as the guided images and decision maps are used as input images in JBF. After filtering operation by JBF, the output images are treated as weight maps. The sub-band coefficients are fused by weighted average algorithm. Finally, the fused sub-bands are composed into a fused image by the inverse NSCT. The experiments on the IFR and VI image are carried out. For the assessment of fusion results, subjective and objective assessment methods are adopted. The results show that the proposed methodology can get better performance than some classical fusion method existed in the published literatures.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"4 1","pages":"1-6"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"78917270","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302251
Juan Gao, Yunsong Qi, Changhui Jiang, N. Zhang, Zhanli Hu
Accurate geometric parameter information for cone-beam CT (CBCT) systems is crucial for high-quality image reconstruction. To validate the performance of the algorithms for geometric parameter extraction and projection matrix computation, this paper presents test calibration methods based on a computer simulation. We simulate the projection operation on a calibration phantom using Visual C++ and obtain the center of projection images through an approach based on least squares and genetic algorithm using Matlab programs. To verify the performance of the presented geometric calibration algorithm for projection matrix computation and geometric parameter extraction, CBCT consisting of a flat-panel detector is simulated and the un-calibration reconstructed image is compared with the reconstructed images of the calibration method in this paper. The extracted geometric parameters from the calculated projection matrix are very close to the input values. Compared with the uncorrected reconstructed image, the corrected reconstructed image significantly reduces many artifacts. Experimental results reveal that the presented method is robust and accurate, and can suppress undesirable artifacts of reconstructed images which caused by misaligned scanner geometry.
{"title":"A projection matrix-based geometric calibration algorithm in CBCT system","authors":"Juan Gao, Yunsong Qi, Changhui Jiang, N. Zhang, Zhanli Hu","doi":"10.1109/CISP-BMEI.2017.8302251","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302251","url":null,"abstract":"Accurate geometric parameter information for cone-beam CT (CBCT) systems is crucial for high-quality image reconstruction. To validate the performance of the algorithms for geometric parameter extraction and projection matrix computation, this paper presents test calibration methods based on a computer simulation. We simulate the projection operation on a calibration phantom using Visual C++ and obtain the center of projection images through an approach based on least squares and genetic algorithm using Matlab programs. To verify the performance of the presented geometric calibration algorithm for projection matrix computation and geometric parameter extraction, CBCT consisting of a flat-panel detector is simulated and the un-calibration reconstructed image is compared with the reconstructed images of the calibration method in this paper. The extracted geometric parameters from the calculated projection matrix are very close to the input values. Compared with the uncorrected reconstructed image, the corrected reconstructed image significantly reduces many artifacts. Experimental results reveal that the presented method is robust and accurate, and can suppress undesirable artifacts of reconstructed images which caused by misaligned scanner geometry.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"314 1","pages":"1-5"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"78938248","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2017-10-01DOI: 10.1109/CISP-BMEI.2017.8302311
Wenqiao Sun
Named entity recognition(NER) has been studied for a long time as more and more researches about the embedding, neural network model and some others systems like Language Model have developed quickly. However, as these systems rely heavily on domain-specific knowledge and it can hardly acquires much data about Chinese postal addresses, Chinese Named entity recognition(CNER) task on postal address has developed slowly. In this paper, we use a modified Conditional Random Field(CRF) model to solve a CNER task on a postal address corpus. Since there has little data about Chinese postal addresses and parts of which are incomplete sentences, we utilize the known, useful, clearer semantics words and sentences to our model as the additional features. We make three experiments to evaluate our system which obtains good performance and it shows that our modified algorithm performs better than other traditional algorithms when processing postal addresses.
{"title":"Chinese named entity recognition using modified conditional random field on postal address","authors":"Wenqiao Sun","doi":"10.1109/CISP-BMEI.2017.8302311","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302311","url":null,"abstract":"Named entity recognition(NER) has been studied for a long time as more and more researches about the embedding, neural network model and some others systems like Language Model have developed quickly. However, as these systems rely heavily on domain-specific knowledge and it can hardly acquires much data about Chinese postal addresses, Chinese Named entity recognition(CNER) task on postal address has developed slowly. In this paper, we use a modified Conditional Random Field(CRF) model to solve a CNER task on a postal address corpus. Since there has little data about Chinese postal addresses and parts of which are incomplete sentences, we utilize the known, useful, clearer semantics words and sentences to our model as the additional features. We make three experiments to evaluate our system which obtains good performance and it shows that our modified algorithm performs better than other traditional algorithms when processing postal addresses.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"22 1","pages":"1-6"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"79149410","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Exon is an important functional region of eukaryotic DNA sequence. Prediction of exons can help to understand the structure and function of protein. However, the issue of finding an efficient technique to detect the numbers and locations of short coding sequences automatically is an unsolved problem. In this work, a short exon prediction method based on multiscale products in B-spline wavelet domain is proposed. The proposed wavelet denoising and multiscale products-based technique (WDMP) for short exons prediction have the following three features. (1) A wavelet package denoising method is applied to smooth the DNA numerical sequences. (2) A new B-spline wavelet function is designed to extract the exon features in multiscale domain, so the effect of window length is avoided. In addition, this wavelet has a higher degree of freedom for curve design. (3) We multiply the adjacent coefficients to exploit the high inter-scale correlation of the exon data, while these correlation features are used to separate the exon signals from background noise. Compared with four well-known model-independent methods, case studies demonstrate that the proposed WDMP method helps to improve the prediction accuracy of short exons significantly.
{"title":"Improved prediction of short exons via multiscale products","authors":"Guishan Zhang, Xiaolei Zhang, Guocheng Pan, Yangjiang Yu, Yaowen Chen","doi":"10.1109/CISP-BMEI.2017.8302225","DOIUrl":"https://doi.org/10.1109/CISP-BMEI.2017.8302225","url":null,"abstract":"Exon is an important functional region of eukaryotic DNA sequence. Prediction of exons can help to understand the structure and function of protein. However, the issue of finding an efficient technique to detect the numbers and locations of short coding sequences automatically is an unsolved problem. In this work, a short exon prediction method based on multiscale products in B-spline wavelet domain is proposed. The proposed wavelet denoising and multiscale products-based technique (WDMP) for short exons prediction have the following three features. (1) A wavelet package denoising method is applied to smooth the DNA numerical sequences. (2) A new B-spline wavelet function is designed to extract the exon features in multiscale domain, so the effect of window length is avoided. In addition, this wavelet has a higher degree of freedom for curve design. (3) We multiply the adjacent coefficients to exploit the high inter-scale correlation of the exon data, while these correlation features are used to separate the exon signals from background noise. Compared with four well-known model-independent methods, case studies demonstrate that the proposed WDMP method helps to improve the prediction accuracy of short exons significantly.","PeriodicalId":6474,"journal":{"name":"2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)","volume":"39 1","pages":"1-6"},"PeriodicalIF":0.0,"publicationDate":"2017-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"79223982","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：