Pub Date : 2021-07-16DOI: 10.1109/PRML52754.2021.9520689
Guyu Fang, Hongtao Lu
Most of existing unsupervised person re-identification methods focus on cross-domain adaptation. In order to further relieve the dependence on manual labels, we propose a camera-aware three-stage method for fully unsupervised person re-identification which only requires the unlabeled target dataset. We exploit camera labels and divide the learning process into three relatively easy sub-tasks: initialization by instance discrimination, intra-camera learning and inter-camera learning. The first stage regards each person image as an instance and tries to distinguish each image. The second stage performs intra-camera clustering while the last stage performs clustering and training on the whole dataset. These three stages share the backbone network. Finally, our method substantially boosts the performance stage by stage without any manual ID annotation. We conduct extensive experiments on three large-scale image-based datasets, including Market-1501, DukeMTMC-reID and MSMT17. The results demonstrate that our method achieves the state-of-the-art performance.
{"title":"A Camera-Aware Three-Stage Method for Fully Unsupervised Person Re-identification","authors":"Guyu Fang, Hongtao Lu","doi":"10.1109/PRML52754.2021.9520689","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520689","url":null,"abstract":"Most of existing unsupervised person re-identification methods focus on cross-domain adaptation. In order to further relieve the dependence on manual labels, we propose a camera-aware three-stage method for fully unsupervised person re-identification which only requires the unlabeled target dataset. We exploit camera labels and divide the learning process into three relatively easy sub-tasks: initialization by instance discrimination, intra-camera learning and inter-camera learning. The first stage regards each person image as an instance and tries to distinguish each image. The second stage performs intra-camera clustering while the last stage performs clustering and training on the whole dataset. These three stages share the backbone network. Finally, our method substantially boosts the performance stage by stage without any manual ID annotation. We conduct extensive experiments on three large-scale image-based datasets, including Market-1501, DukeMTMC-reID and MSMT17. The results demonstrate that our method achieves the state-of-the-art performance.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"22 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123951523","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520714
Zhensong Li, Jie Zhu, Hong Cao
In the field of natural language processing, pre-training can effectively improve the performance of downstream tasks. In recent years, pre-training has been continuously developed in Tibetan NLP. We built three pre-trained models of Tibetan Word2Vec, Tibetan ELMo, and Tibetan ALBERT, and applied them to the two downstream tasks of Tibetan text classification and Tibetan part-of-speech tagging. Comparing them with the baseline models of these two downstream tasks, it is found that the performance of the downstream tasks using the pre-training is significantly better than the baseline model. The three pre-trained models have also brought a gradual improvement in performance for Tibetan downstream tasks.
{"title":"Research on Pre-training of Tibetan Natural Language Processing","authors":"Zhensong Li, Jie Zhu, Hong Cao","doi":"10.1109/PRML52754.2021.9520714","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520714","url":null,"abstract":"In the field of natural language processing, pre-training can effectively improve the performance of downstream tasks. In recent years, pre-training has been continuously developed in Tibetan NLP. We built three pre-trained models of Tibetan Word2Vec, Tibetan ELMo, and Tibetan ALBERT, and applied them to the two downstream tasks of Tibetan text classification and Tibetan part-of-speech tagging. Comparing them with the baseline models of these two downstream tasks, it is found that the performance of the downstream tasks using the pre-training is significantly better than the baseline model. The three pre-trained models have also brought a gradual improvement in performance for Tibetan downstream tasks.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"43 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125382124","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520711
Meng Yan, L. Li, Haochuan Chen
In VideoSAR circle trace imaging mode, the energy of moving target is defocused and shifted. However, due to the occlusion of target height, there is shadow in its real position, which represents the lack of energy. In addition, there is a strong correlation between adjacent frames of VideoSAR image sequence, and the shadow also moves with the movement of the target. Based on this property, a new method for moving object detection in VideoSAR image sequences is proposed. This method is based on Gaussian mixture model. Firstly, it preprocesses the image sequence, uses sift + RANSAC algorithm and median filter processing, then uses Otsu threshold segmentation algorithm to transform the image into binary image, uses Gaussian mixture model to detect moving objects, and finally carries out morphological processing. Using VideoSAR image sequence of Sandia National Laboratory, the moving target can be detected effectively.
{"title":"A VideoSAR Moving Target Detection Method Based on GMM","authors":"Meng Yan, L. Li, Haochuan Chen","doi":"10.1109/PRML52754.2021.9520711","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520711","url":null,"abstract":"In VideoSAR circle trace imaging mode, the energy of moving target is defocused and shifted. However, due to the occlusion of target height, there is shadow in its real position, which represents the lack of energy. In addition, there is a strong correlation between adjacent frames of VideoSAR image sequence, and the shadow also moves with the movement of the target. Based on this property, a new method for moving object detection in VideoSAR image sequences is proposed. This method is based on Gaussian mixture model. Firstly, it preprocesses the image sequence, uses sift + RANSAC algorithm and median filter processing, then uses Otsu threshold segmentation algorithm to transform the image into binary image, uses Gaussian mixture model to detect moving objects, and finally carries out morphological processing. Using VideoSAR image sequence of Sandia National Laboratory, the moving target can be detected effectively.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125114429","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520717
Shuangjiang Du, Baofu Zhang, Pin Zhang, Peng Xiang
The regression loss function is a key factor in the training and optimization process of object detection. The current mainstream regression loss functions are Ln norm loss, IOU loss and CIOU loss. This paper proposes the Scale-Sensitive IOU(SIOU) loss, a new loss function different from the above all, which could solve the issues that the current loss functions cannot distinguish the two bounding boxes in some special cases when the target area scales in one image vary greatly during training process, thereby leading to the improper regression loss calculation and the slowing down of the optimization. An area scale regulating factor Y is added on the basis of CIOU loss to adjust the loss values of the bounding boxes, which could distinguish all the boxes quantitatively in theory thus gets a faster converging speed and better optimization. Through analysis and simulation comparison among the several loss functions, the superiority of SIOU loss is verified. Furthermore, by incorporating SIOU loss into YOLO v4, Faster R-CNN and SSD on the two mainstream aerial remote sensing datasets, i.e., DIOR and NWPU VHR-10, the detection precisions improve by 10.2% than IOU loss and 2.8% than CIOU loss respectively.
{"title":"An Improved Bounding Box Regression Loss Function Based on CIOU Loss for Multi-scale Object Detection","authors":"Shuangjiang Du, Baofu Zhang, Pin Zhang, Peng Xiang","doi":"10.1109/PRML52754.2021.9520717","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520717","url":null,"abstract":"The regression loss function is a key factor in the training and optimization process of object detection. The current mainstream regression loss functions are Ln norm loss, IOU loss and CIOU loss. This paper proposes the Scale-Sensitive IOU(SIOU) loss, a new loss function different from the above all, which could solve the issues that the current loss functions cannot distinguish the two bounding boxes in some special cases when the target area scales in one image vary greatly during training process, thereby leading to the improper regression loss calculation and the slowing down of the optimization. An area scale regulating factor Y is added on the basis of CIOU loss to adjust the loss values of the bounding boxes, which could distinguish all the boxes quantitatively in theory thus gets a faster converging speed and better optimization. Through analysis and simulation comparison among the several loss functions, the superiority of SIOU loss is verified. Furthermore, by incorporating SIOU loss into YOLO v4, Faster R-CNN and SSD on the two mainstream aerial remote sensing datasets, i.e., DIOR and NWPU VHR-10, the detection precisions improve by 10.2% than IOU loss and 2.8% than CIOU loss respectively.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"12 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124215119","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}
The negFIN is the current state-of-the art algorithm for frequent itemsets mining. It employs a novel BMC (bitmap code) encoding model for nodes in a prefix tree based on the bitmap representation of sets. The encoding of each node is a binary number of which bit number is the number of frequent items, and is stored in the form of decimal integer number. The key operations of negFIN are all performed based on the bitwise operation of the encoding. The main problem of BMC is that the maximal bit number of the data type which is used to store the decimal integer number in current general compiling systems is 64, so if the number of frequent items exceeds 64, the encoding cannot work effectively. In this work, we propose B-BMC (block bitmap code) encoding model, a more efficient encoding model. The B-BMC is a dividing of BMC based on the block size in essential. For facilitating the work of B-BMC, the B-BMC tree and TNC(terminal node code) table are devised as an alternative to the BMC tree of negFIN. Based on these two structures, we present an efficient algorithm called INFIN (improved negFIN) to mining frequent itemsets. Our experiments illustrate that the B-BMC can overcome the drawback of BMC, and the INFIN is the most efficient one in time and space when the block size takes value 64 on condition that the number of frequent items exceeds 64.
{"title":"INFIN: An Efficient Algorithm for Fast Mining Frequent Itemsets","authors":"Shaopeng Wang, Yufei Wang, Chunkai Feng, ChaoYu Niu","doi":"10.1109/PRML52754.2021.9520736","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520736","url":null,"abstract":"The negFIN is the current state-of-the art algorithm for frequent itemsets mining. It employs a novel BMC (bitmap code) encoding model for nodes in a prefix tree based on the bitmap representation of sets. The encoding of each node is a binary number of which bit number is the number of frequent items, and is stored in the form of decimal integer number. The key operations of negFIN are all performed based on the bitwise operation of the encoding. The main problem of BMC is that the maximal bit number of the data type which is used to store the decimal integer number in current general compiling systems is 64, so if the number of frequent items exceeds 64, the encoding cannot work effectively. In this work, we propose B-BMC (block bitmap code) encoding model, a more efficient encoding model. The B-BMC is a dividing of BMC based on the block size in essential. For facilitating the work of B-BMC, the B-BMC tree and TNC(terminal node code) table are devised as an alternative to the BMC tree of negFIN. Based on these two structures, we present an efficient algorithm called INFIN (improved negFIN) to mining frequent itemsets. Our experiments illustrate that the B-BMC can overcome the drawback of BMC, and the INFIN is the most efficient one in time and space when the block size takes value 64 on condition that the number of frequent items exceeds 64.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"22 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131938185","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520744
Liang Ziyao, Yu Yong, Zhang Bin
The radar echoes of the corner reflector array and the ship target are very similar, and the existing algorithms are difficult to identify them effectively in time, frequency and spatial domain. Aiming at the problem that the terminal guidance radar of anti-ship missile can’t detect and track the real target effectively under the deception jamming of corner reflector array, this paper designs a countermeasure method based on Pauli polarization decomposition and BP neural network. Firstly, the Pauli polarization decomposition of the full polarization scattering matrix of the target measured in the fixed angle window is used to obtain four normalized coefficients and form the eigenvector, and the differences between the ship target and the corner reflector are analyzed. Then, the BP neural network model is trained and optimized as the training sample. The simulation and test results show that the feature vectors can distinguish the two kinds of targets, and the trained network can identify the ship and the corner reflector Array effectively, and the overall success rate is close to 97%.
{"title":"Anti-Corner Reflector Array Method Based on Pauli Polarization Decomposition and BP Neural Network","authors":"Liang Ziyao, Yu Yong, Zhang Bin","doi":"10.1109/PRML52754.2021.9520744","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520744","url":null,"abstract":"The radar echoes of the corner reflector array and the ship target are very similar, and the existing algorithms are difficult to identify them effectively in time, frequency and spatial domain. Aiming at the problem that the terminal guidance radar of anti-ship missile can’t detect and track the real target effectively under the deception jamming of corner reflector array, this paper designs a countermeasure method based on Pauli polarization decomposition and BP neural network. Firstly, the Pauli polarization decomposition of the full polarization scattering matrix of the target measured in the fixed angle window is used to obtain four normalized coefficients and form the eigenvector, and the differences between the ship target and the corner reflector are analyzed. Then, the BP neural network model is trained and optimized as the training sample. The simulation and test results show that the feature vectors can distinguish the two kinds of targets, and the trained network can identify the ship and the corner reflector Array effectively, and the overall success rate is close to 97%.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"202 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134480157","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520720
P. Nanthakumaran, L. Liyanage
An understanding on the exposure to environmental factors aggravating global disease burden can aid mitigating it. Generally, a class of generalized linear models and generalized additive models are used in predicting disease burden whereas, tree-based models are underused. The objective of this paper is to evaluate the performance of different tree-based models namely decision tree, random forest, gradient boosted tree and stochastic gradient boosted trees in predicting asthma attack based on short-term exposure to environmental factors and to examine the environmental factors triggering asthma attack. A sample of patients during 2013 - 2015 from different parts of Victoria was considered. The study area for the considered study period had reasonably good air quality and relatively humid environment. The tree-based models were tuned using random grid search optimization with bootstrapping to address over-fitting. The models considered performed well in predicting asthma attacks in terms of area under the receiver operating curve (ROC AUC) (>0.82). All the gradient boosted trees (accuracy = 76%; recall = 63%; F2-score = 64%) showed better overall prediction whereas decision tree (accuracy = 71%; recall = 75%; F2-score = 71%) outperformed other models in identifying the positive cases. Tree-based models revealed that O3 exposure consistently influence Asthma. Further, decision tree revealed O3 exposure < 13 ppb or with high O3 exposure >= 13 ppb, and with [SO2 exposure < 0.5 ppb and maximum wind speed > 5.4. km/hr.] influenced Asthma. In addition, relative humidity and exposure to CO were also detected in other tree-based models as relevant predictors triggering asthma attacks.
{"title":"Tree-Based Models Using Random Grid Search Optimization for Disease Classification Based on Environmental Factors: A Case Study on Asthma Hospitalizations","authors":"P. Nanthakumaran, L. Liyanage","doi":"10.1109/PRML52754.2021.9520720","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520720","url":null,"abstract":"An understanding on the exposure to environmental factors aggravating global disease burden can aid mitigating it. Generally, a class of generalized linear models and generalized additive models are used in predicting disease burden whereas, tree-based models are underused. The objective of this paper is to evaluate the performance of different tree-based models namely decision tree, random forest, gradient boosted tree and stochastic gradient boosted trees in predicting asthma attack based on short-term exposure to environmental factors and to examine the environmental factors triggering asthma attack. A sample of patients during 2013 - 2015 from different parts of Victoria was considered. The study area for the considered study period had reasonably good air quality and relatively humid environment. The tree-based models were tuned using random grid search optimization with bootstrapping to address over-fitting. The models considered performed well in predicting asthma attacks in terms of area under the receiver operating curve (ROC AUC) (>0.82). All the gradient boosted trees (accuracy = 76%; recall = 63%; F2-score = 64%) showed better overall prediction whereas decision tree (accuracy = 71%; recall = 75%; F2-score = 71%) outperformed other models in identifying the positive cases. Tree-based models revealed that O3 exposure consistently influence Asthma. Further, decision tree revealed O3 exposure < 13 ppb or with high O3 exposure >= 13 ppb, and with [SO2 exposure < 0.5 ppb and maximum wind speed > 5.4. km/hr.] influenced Asthma. In addition, relative humidity and exposure to CO were also detected in other tree-based models as relevant predictors triggering asthma attacks.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131194625","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520724
Z. Ding, Haibin Xie, Peng Li
The performance of the clustering algorithms depends mainly on the setting of artificial parameter values which is usually difficult in practical application. In addition, the dataset is usually incremental, and the clustering algorithm applied to the static dataset cannot develop with the change of the dataset. If new sample points are added, algorithm parameters need to be readjusted to cluster again, leading to a great time cost. This paper proposed an evolutionary parameter-free clustering algorithm (EPFC) for the above problems, which imitates the human clustering mechanism of objective things. EPFC algorithm takes the average distance between each sample and its nearest neighbour sample as the threshold value to judge whether the sample can be grouped into one cluster. The threshold value is adaptively updated without setting an artificially parameter value as the samples increase. A large number of experiments on benchmark datasets show that EPFC is effective on datasets with different characteristics, and the algorithm has strong robustness.
{"title":"Evolutionary Parameter-Free Clustering Algorithm","authors":"Z. Ding, Haibin Xie, Peng Li","doi":"10.1109/PRML52754.2021.9520724","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520724","url":null,"abstract":"The performance of the clustering algorithms depends mainly on the setting of artificial parameter values which is usually difficult in practical application. In addition, the dataset is usually incremental, and the clustering algorithm applied to the static dataset cannot develop with the change of the dataset. If new sample points are added, algorithm parameters need to be readjusted to cluster again, leading to a great time cost. This paper proposed an evolutionary parameter-free clustering algorithm (EPFC) for the above problems, which imitates the human clustering mechanism of objective things. EPFC algorithm takes the average distance between each sample and its nearest neighbour sample as the threshold value to judge whether the sample can be grouped into one cluster. The threshold value is adaptively updated without setting an artificially parameter value as the samples increase. A large number of experiments on benchmark datasets show that EPFC is effective on datasets with different characteristics, and the algorithm has strong robustness.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"37 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133288924","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520688
Ya Zhang, Shuai Qin
As a basis task in the field of Natural Language Processing (NLP), relation extraction task aims to extract the relation between two entities in a text. Most existing models rely on a single semantic feature of the sentence for relation classification. In this paper, we present MGGCM model, a novel neural network to classify the relation of two entities in a sentence. Our neural architecture leverages two distinct graphs which are the dependency tree path and the relation-entity graph respectively. In this model, we integrate both semantic features and structural features to enhance the performance of relation extraction model. We encode the sentence through BiLSTM, obtain its structural features by GCN, and pay more attention to the entity information which is related to the target entity pair, and finally fuse the features to obtain the classification results. We test our model on the SemEval 2010 relation classification task, and achieve an F1-score of 85.7%, higher than competing methods in literature.
{"title":"Improving Relation Classification with Multi-graph GCN","authors":"Ya Zhang, Shuai Qin","doi":"10.1109/PRML52754.2021.9520688","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520688","url":null,"abstract":"As a basis task in the field of Natural Language Processing (NLP), relation extraction task aims to extract the relation between two entities in a text. Most existing models rely on a single semantic feature of the sentence for relation classification. In this paper, we present MGGCM model, a novel neural network to classify the relation of two entities in a sentence. Our neural architecture leverages two distinct graphs which are the dependency tree path and the relation-entity graph respectively. In this model, we integrate both semantic features and structural features to enhance the performance of relation extraction model. We encode the sentence through BiLSTM, obtain its structural features by GCN, and pay more attention to the entity information which is related to the target entity pair, and finally fuse the features to obtain the classification results. We test our model on the SemEval 2010 relation classification task, and achieve an F1-score of 85.7%, higher than competing methods in literature.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"5 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121974088","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 : 2021-07-16DOI: 10.1109/PRML52754.2021.9520728
Gui-Hong Shi, Jiezhong Huang, Junhua Zhang, Guoqin Tan, Gaoli Sang
Accuracy target detection can benefit many target detection applications. The latest YOLOv5 method has faster detection speed and better accuracy in target detection. However, there are still insufficient on bounding box positioning and it is difficult to distinguish overlapping objects. This paper proposes an improved Attention-YOLO v5, which adds channel attention and spatial attention mechanisms to the feature extraction. Furthermore, a squeeze and excitation(SE) module is applied to improve the resolution of the input image. Experiments on two public datasets show that our proposed method effectively reduces the positioning error of the bounding box and improves the detection accuracy. The accuracy on INRIA and PnPLO datasets are 97.9% and 96.2%.
{"title":"Combined Channel and Spatial Attention for YOLOv5 during Target Detection","authors":"Gui-Hong Shi, Jiezhong Huang, Junhua Zhang, Guoqin Tan, Gaoli Sang","doi":"10.1109/PRML52754.2021.9520728","DOIUrl":"https://doi.org/10.1109/PRML52754.2021.9520728","url":null,"abstract":"Accuracy target detection can benefit many target detection applications. The latest YOLOv5 method has faster detection speed and better accuracy in target detection. However, there are still insufficient on bounding box positioning and it is difficult to distinguish overlapping objects. This paper proposes an improved Attention-YOLO v5, which adds channel attention and spatial attention mechanisms to the feature extraction. Furthermore, a squeeze and excitation(SE) module is applied to improve the resolution of the input image. Experiments on two public datasets show that our proposed method effectively reduces the positioning error of the bounding box and improves the detection accuracy. The accuracy on INRIA and PnPLO datasets are 97.9% and 96.2%.","PeriodicalId":429603,"journal":{"name":"2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML)","volume":"15 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127900387","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：