We propose an efficient hybrid method that combines neural network and particle swarm optimization algorithm to optimize the performance of backward multi-pumped Raman fiber amplifiers. We use a neural network to inverse system design Raman fiber amplifier by learning the nonlinear mapping relationship between pump light and the output gain. To obtain a flat gain spectrum, the particle swarm optimization algorithm is used to search for the optimal pump slight parameter configuration. The results show that when the designed Raman amplifier is oriented toward C+L band signal optical amplification, the error between the target gain value and the actual gain value is less than 0.47 dB, the output gain after optimization is 17.96dB, and the gain flatness is 0.44dB.
{"title":"Research on Raman fiber amplifier using neural network combining PSO algorithm","authors":"J. Gong, Jiaojiao Lu, Ruijie Gao","doi":"10.1145/3573942.3573956","DOIUrl":"https://doi.org/10.1145/3573942.3573956","url":null,"abstract":"We propose an efficient hybrid method that combines neural network and particle swarm optimization algorithm to optimize the performance of backward multi-pumped Raman fiber amplifiers. We use a neural network to inverse system design Raman fiber amplifier by learning the nonlinear mapping relationship between pump light and the output gain. To obtain a flat gain spectrum, the particle swarm optimization algorithm is used to search for the optimal pump slight parameter configuration. The results show that when the designed Raman amplifier is oriented toward C+L band signal optical amplification, the error between the target gain value and the actual gain value is less than 0.47 dB, the output gain after optimization is 17.96dB, and the gain flatness is 0.44dB.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"43 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127545516","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}
Aiming at the problem that the maximum number of continuous uniform array elements of the virtual array extended by the coprime array algorithm is small and the degree of freedom is still low. A matrix reconstruction DOA estimation algorithm based on virtual array interpolation is proposed. Firstly, the general coprime array is improved by optimizing the array layout to form a new array, and the new array is derived from a non-uniform virtual array, which increases the number of array elements and improves the degree of freedom; secondly, the idea of virtual array interpolation is used to fill the holes in the virtual domain A uniform linear virtual array is constructed, and finally the DOA is estimated by optimizing the design through atomic norm minimization and sparse reconstruction of the covariance matrix. The algorithm improves the degree of freedom of the array and makes full use of the information in the virtual array. The simulation results show the effectiveness of the new array algorithm.
{"title":"DOA Estimation of Shifted Coprime Array Based on Covariance Matrix Reconstruction","authors":"Wei Yang, Dongming Xu, Jiaqi Xue","doi":"10.1145/3573942.3574112","DOIUrl":"https://doi.org/10.1145/3573942.3574112","url":null,"abstract":"Aiming at the problem that the maximum number of continuous uniform array elements of the virtual array extended by the coprime array algorithm is small and the degree of freedom is still low. A matrix reconstruction DOA estimation algorithm based on virtual array interpolation is proposed. Firstly, the general coprime array is improved by optimizing the array layout to form a new array, and the new array is derived from a non-uniform virtual array, which increases the number of array elements and improves the degree of freedom; secondly, the idea of virtual array interpolation is used to fill the holes in the virtual domain A uniform linear virtual array is constructed, and finally the DOA is estimated by optimizing the design through atomic norm minimization and sparse reconstruction of the covariance matrix. The algorithm improves the degree of freedom of the array and makes full use of the information in the virtual array. The simulation results show the effectiveness of the new array algorithm.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"13 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127733675","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}
Aiming at the problem that the laser imaging fuze needs to process a large amount of data and the algorithm is complex, so it is difficult to eliminate the cloud interference quickly and accurately, a target recognition method based on the combination of improved Harris corner detection algorithm and rectangularity is proposed. Firstly, B-spline function is used to replace the Gaussian window function in the original corner detection algorithm for smoothing filtering; Secondly, the gray value of the central pixel is compared with its 8 neighborhood, and the diagonal points are pre screened. After eliminating the pseudo corners, the corners are determined by using the improved corner response function and non maximum suppression; Finally, count the number of corners and calculate the rectangularity of the corner area, take the number of corners and rectangularity as the feature vector, and select the linear analysis method with the highest accuracy and the fastest reasoning speed for classification and recognition. The experimental results show that the accuracy of the target recognition method can reach 95.02%. The target recognition method proposed in this paper can quickly and accurately distinguish fighter from cloud.
{"title":"Target Recognition of Laser Imaging Fuze Based on Corner Features","authors":"Lina Liu, W. He","doi":"10.1145/3573942.3574081","DOIUrl":"https://doi.org/10.1145/3573942.3574081","url":null,"abstract":"Aiming at the problem that the laser imaging fuze needs to process a large amount of data and the algorithm is complex, so it is difficult to eliminate the cloud interference quickly and accurately, a target recognition method based on the combination of improved Harris corner detection algorithm and rectangularity is proposed. Firstly, B-spline function is used to replace the Gaussian window function in the original corner detection algorithm for smoothing filtering; Secondly, the gray value of the central pixel is compared with its 8 neighborhood, and the diagonal points are pre screened. After eliminating the pseudo corners, the corners are determined by using the improved corner response function and non maximum suppression; Finally, count the number of corners and calculate the rectangularity of the corner area, take the number of corners and rectangularity as the feature vector, and select the linear analysis method with the highest accuracy and the fastest reasoning speed for classification and recognition. The experimental results show that the accuracy of the target recognition method can reach 95.02%. The target recognition method proposed in this paper can quickly and accurately distinguish fighter from cloud.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121561234","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}
Sheping Zhai, Wenqing Zhang, Dabao Cheng, Xiaoxia Bai
Extracting and representing text features is the most important part of text classification. Aiming at the problem of incomplete feature extraction in traditional text classification methods, a text classification model based on graph convolution neural network and attention mechanism is proposed. Firstly, the text is input into BERT (Bi-directional Encoder Representations from Transformers) model to obtain the word vector representation, the context semantic information of the given text is learned by the BiGRU (Bi-directional Gated Recurrent Unit), and the important information is screened by attention mechanism and used as node features. Secondly, the dependency syntax diagram and the corresponding adjacency matrix of the input text are constructed. Thirdly, the GCN (Graph Convolution Neural Network) is used to learn the node features and adjacency matrix. Finally, the obtained text features are input into the classifier for text classification. Experiments on two datasets show that the proposed model achieves a good classification effect, and better accuracy is achieved in comparison with baseline models.
文本特征的提取和表示是文本分类的重要组成部分。针对传统文本分类方法中特征提取不完全的问题，提出了一种基于图卷积神经网络和注意机制的文本分类模型。首先，将文本输入到BERT (Bi-directional Encoder Representations from Transformers)模型中获得词向量表示，通过双向门控循环单元(Bi-directional Gated Recurrent Unit)学习给定文本的上下文语义信息，通过注意机制筛选重要信息作为节点特征;其次，构造输入文本的依赖句法图和相应的邻接矩阵;第三，利用GCN(图卷积神经网络)学习节点特征和邻接矩阵。最后，将得到的文本特征输入到分类器中进行文本分类。在两个数据集上的实验表明，该模型取得了良好的分类效果，与基线模型相比，准确率更高。
{"title":"Text Classification Based on Graph Convolution Neural Network and Attention Mechanism","authors":"Sheping Zhai, Wenqing Zhang, Dabao Cheng, Xiaoxia Bai","doi":"10.1145/3573942.3573963","DOIUrl":"https://doi.org/10.1145/3573942.3573963","url":null,"abstract":"Extracting and representing text features is the most important part of text classification. Aiming at the problem of incomplete feature extraction in traditional text classification methods, a text classification model based on graph convolution neural network and attention mechanism is proposed. Firstly, the text is input into BERT (Bi-directional Encoder Representations from Transformers) model to obtain the word vector representation, the context semantic information of the given text is learned by the BiGRU (Bi-directional Gated Recurrent Unit), and the important information is screened by attention mechanism and used as node features. Secondly, the dependency syntax diagram and the corresponding adjacency matrix of the input text are constructed. Thirdly, the GCN (Graph Convolution Neural Network) is used to learn the node features and adjacency matrix. Finally, the obtained text features are input into the classifier for text classification. Experiments on two datasets show that the proposed model achieves a good classification effect, and better accuracy is achieved in comparison with baseline models.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117086549","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}
Zongyu Xu* School of Computer Science and Technology, Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi’an University of Posts & Telecommunications, Xi’an, Shaanxi, 710121, China xuzongyu@stu.xupt.edu.cn Xuebin Xu School of Computer Science and Technology, Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi’an University of Posts & Telecommunications, Xi’an, Shaanxi, 710121, China xuxuebin@xupt.edu.cn Zihao Huang School of Computer Science and Technology, Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi’an University of Posts & Telecommunications, Xi’an, Shaanxi, 710121, China 1277430572@qq.com
{"title":"The Research of Retinopathy Image Recognition Method Based on Vit","authors":"Zongyu Xu, Xuebin Xu, Zihao Huang","doi":"10.1145/3573942.3574083","DOIUrl":"https://doi.org/10.1145/3573942.3574083","url":null,"abstract":"Zongyu Xu* School of Computer Science and Technology, Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi’an University of Posts & Telecommunications, Xi’an, Shaanxi, 710121, China xuzongyu@stu.xupt.edu.cn Xuebin Xu School of Computer Science and Technology, Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi’an University of Posts & Telecommunications, Xi’an, Shaanxi, 710121, China xuxuebin@xupt.edu.cn Zihao Huang School of Computer Science and Technology, Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi’an University of Posts & Telecommunications, Xi’an, Shaanxi, 710121, China 1277430572@qq.com","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115537952","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}
Traditional convolution and pooling operations in the previous semantic segmentation methods will cause the loss of feature information due to limited receptive field size. They are insufficient to support an accurate image prediction result. To solve this problem, Firstly, we design a Second-Order Encoder to enlarge the feature receptive field and capture more semantic context information. Secondly, we design a Restore Detail Decoder to focus on processing the spatial detail information and refining the object edges. The experiments verify the effectiveness of the proposed approach. The results show that our method achieves competitive performance on two datasets, including PASCAL VOC2012 and Cityscapes with the mIoU of 80.13% and 76.31%, respectively.
{"title":"SECOND-Order Encoder and Restore Detail Decoder Network for Image Semantic Segmentation","authors":"Nan Dai, Zhiqiang Hou, Minjie Cheng","doi":"10.1145/3573942.3574045","DOIUrl":"https://doi.org/10.1145/3573942.3574045","url":null,"abstract":"Traditional convolution and pooling operations in the previous semantic segmentation methods will cause the loss of feature information due to limited receptive field size. They are insufficient to support an accurate image prediction result. To solve this problem, Firstly, we design a Second-Order Encoder to enlarge the feature receptive field and capture more semantic context information. Secondly, we design a Restore Detail Decoder to focus on processing the spatial detail information and refining the object edges. The experiments verify the effectiveness of the proposed approach. The results show that our method achieves competitive performance on two datasets, including PASCAL VOC2012 and Cityscapes with the mIoU of 80.13% and 76.31%, respectively.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115846152","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 existing methods for forecasting clothing trends mostly use traditional time series forecasting methods, and the data sources are mostly sale data from e-commerce websites, which have large errors in forecasting accuracy. This paper proposes a new model CNN-BiLSTM-Attention for predicting clothing trends based on social media data. The Geostyle dataset is pre-processed to get the clothing popularity index. First, One-dimensional CNN is used to extract the important features in the clothing popularity index. Second, the BiLSTM is used to make full use of contextual information. Third, adding an Attention mechanism to the output can highlight relevant information, suppress irrelevant information, and significantly improve prediction accuracy. The experimental results show that our method is significantly better than other traditional time series forecasting methods and existing deep learning methods when applied to apparel trend forecasting.
{"title":"Research on Apparel Trend Prediction Based on CNN-BiLSTM-Attention Model","authors":"Chunfa Zhang, Ning Chen, Shu-xu Zhao","doi":"10.1145/3573942.3573953","DOIUrl":"https://doi.org/10.1145/3573942.3573953","url":null,"abstract":"The existing methods for forecasting clothing trends mostly use traditional time series forecasting methods, and the data sources are mostly sale data from e-commerce websites, which have large errors in forecasting accuracy. This paper proposes a new model CNN-BiLSTM-Attention for predicting clothing trends based on social media data. The Geostyle dataset is pre-processed to get the clothing popularity index. First, One-dimensional CNN is used to extract the important features in the clothing popularity index. Second, the BiLSTM is used to make full use of contextual information. Third, adding an Attention mechanism to the output can highlight relevant information, suppress irrelevant information, and significantly improve prediction accuracy. The experimental results show that our method is significantly better than other traditional time series forecasting methods and existing deep learning methods when applied to apparel trend forecasting.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"231 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127009506","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}
Abstract: Diabetes is a common disease that seriously endangers human health, mostly in the middle-aged and the elderly. Predicting the incidence rate of diabetes enables doctors to make a scientific treatment plan in advance, which will significantly improve the cure rate and reduce the incidence rate. Based on this situation, this paper proposes a diabetes prediction model based on ensemble learning, which integrates some classical machine learning algorithms, including Logisticregression, Kneigbors, Decisiontree, GaussianNB, and support vector machine (SVM) The first four low correlation algorithms are constructed as basic learners, and then integrated into meta learner SVM to build an integrated learning model. The advantages of the comprehensive model are evaluated from the following aspects: accuracy, precision, recall rate, AUC, and other evaluation indicators. The experiment was carried out on the Pima Indian diabetes data set (PIDD) published by UCI. First, the XGboost algorithm was used to select the optimal features, and then an integrated learning model was constructed to predict. The experimental results show that the accuracy rate of the integrated learning model is 81.63%, the precision rate is 80%, the recall rate is 80%, and the AUC is 84%. The advantages of the model in accuracy, precision, recall, and AUC are verified. The model will effectively help doctors make more accurate diagnoses and predictions of patients' physical conditions and implement more scientific treatment.
{"title":"A Prediction Model of Diabetes Based on Ensemble Learning","authors":"Lei Qin","doi":"10.1145/3573942.3573949","DOIUrl":"https://doi.org/10.1145/3573942.3573949","url":null,"abstract":"Abstract: Diabetes is a common disease that seriously endangers human health, mostly in the middle-aged and the elderly. Predicting the incidence rate of diabetes enables doctors to make a scientific treatment plan in advance, which will significantly improve the cure rate and reduce the incidence rate. Based on this situation, this paper proposes a diabetes prediction model based on ensemble learning, which integrates some classical machine learning algorithms, including Logisticregression, Kneigbors, Decisiontree, GaussianNB, and support vector machine (SVM) The first four low correlation algorithms are constructed as basic learners, and then integrated into meta learner SVM to build an integrated learning model. The advantages of the comprehensive model are evaluated from the following aspects: accuracy, precision, recall rate, AUC, and other evaluation indicators. The experiment was carried out on the Pima Indian diabetes data set (PIDD) published by UCI. First, the XGboost algorithm was used to select the optimal features, and then an integrated learning model was constructed to predict. The experimental results show that the accuracy rate of the integrated learning model is 81.63%, the precision rate is 80%, the recall rate is 80%, and the AUC is 84%. The advantages of the model in accuracy, precision, recall, and AUC are verified. The model will effectively help doctors make more accurate diagnoses and predictions of patients' physical conditions and implement more scientific treatment.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"15 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121757932","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}
Particle filtering is a standard method for parameter estimation in passive location and has great application value in nonlinear and non-Gaussian systems. The standard particle filter (PF) is prone to the problem of particle weight degradation and particle dilution as the number of iterations increases, which affects the overall performance of the location algorithm. Aiming at this problem, a PF algorithm based on differential squirrel search algorithm (DSSA) optimization is proposed. The particles are divided into optimal individuals, sub-optimal individuals, and ordinary individuals according to the weight of the particles. The low-weight particles are moved closer to the position of the high-weight particles by simulating the predation behavior of squirrels, so that the location information of most particles can be retained. And seasonal monitoring condition is used to avoid the algorithm falling into local optimal. The simulation results show that the improved algorithm has a lower root mean square error (RMSE) than the standard PF algorithm and the improved PF algorithms of other intelligent optimization algorithms under non-Gaussian noise. The improved algorithm can accurately achieve the passive location of the moving target.
{"title":"An Improved Particle Filter Passive Location Method Based on Differential Squirrel Search Algorithm","authors":"Junliang Yang, Ersen Zhang","doi":"10.1145/3573942.3573998","DOIUrl":"https://doi.org/10.1145/3573942.3573998","url":null,"abstract":"Particle filtering is a standard method for parameter estimation in passive location and has great application value in nonlinear and non-Gaussian systems. The standard particle filter (PF) is prone to the problem of particle weight degradation and particle dilution as the number of iterations increases, which affects the overall performance of the location algorithm. Aiming at this problem, a PF algorithm based on differential squirrel search algorithm (DSSA) optimization is proposed. The particles are divided into optimal individuals, sub-optimal individuals, and ordinary individuals according to the weight of the particles. The low-weight particles are moved closer to the position of the high-weight particles by simulating the predation behavior of squirrels, so that the location information of most particles can be retained. And seasonal monitoring condition is used to avoid the algorithm falling into local optimal. The simulation results show that the improved algorithm has a lower root mean square error (RMSE) than the standard PF algorithm and the improved PF algorithms of other intelligent optimization algorithms under non-Gaussian noise. The improved algorithm can accurately achieve the passive location of the moving target.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"2 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114978828","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}
To improve the utilization of convolution features by correlation filter trackers and reduce the influence of interference feature channels on algorithm performance, a correlation filter visual object tracking algorithm based on saliency detection and channel selection is proposed in this paper. Firstly, the HOG features and double-layer convolution features are used to represent the target, and the target salient region mask is obtained by the saliency detection method. Secondly, a channel selection mechanism is designed by using the salient region feature energy and the search region feature energy to remove redundant feature channels containing a large number of background information. Extensive evaluation results obtained on the OTB2015 benchmark demonstrate the effectiveness of the proposed method. The success rate and precision of the proposed algorithm are 67.5% and 91.3%, which are 5.4% and 9.1% higher than the benchmark algorithm BACF, respectively. In addition, according to the experimental results, it can be seen that the proposed tracker has a significant improvement in tracking performance compared with competitors in tracking scenarios with challenges such as deformation, background clutter, and rotation.
{"title":"Correlation Filter Based on Saliency Detection and Channel Selection for Visual Object Tracking","authors":"Sugang Ma, Zhixian Zhao, Lei Zhang, Lei Pu","doi":"10.1145/3573942.3574039","DOIUrl":"https://doi.org/10.1145/3573942.3574039","url":null,"abstract":"To improve the utilization of convolution features by correlation filter trackers and reduce the influence of interference feature channels on algorithm performance, a correlation filter visual object tracking algorithm based on saliency detection and channel selection is proposed in this paper. Firstly, the HOG features and double-layer convolution features are used to represent the target, and the target salient region mask is obtained by the saliency detection method. Secondly, a channel selection mechanism is designed by using the salient region feature energy and the search region feature energy to remove redundant feature channels containing a large number of background information. Extensive evaluation results obtained on the OTB2015 benchmark demonstrate the effectiveness of the proposed method. The success rate and precision of the proposed algorithm are 67.5% and 91.3%, which are 5.4% and 9.1% higher than the benchmark algorithm BACF, respectively. In addition, according to the experimental results, it can be seen that the proposed tracker has a significant improvement in tracking performance compared with competitors in tracking scenarios with challenges such as deformation, background clutter, and rotation.","PeriodicalId":103293,"journal":{"name":"Proceedings of the 2022 5th International Conference on Artificial Intelligence and Pattern Recognition","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115564204","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：