To improve the precision in the later stage of population evolution for multi-objective evolutionary algorithm based on decomposition (MOEA/D), a MOEA/D with integration strategy (MOEA/D-IS) is proposed. The proposed algorithm adopts multiple updating strategies, including a novel first-order differential learning strategy, the individual learning strategy, and the binary and polynomial crossover mutation strategy. The penalty-based boundary intersection approach and Chebyshev approach are used to alternately evaluate individuals. The proposed algorithm and five improved MOEA algorithms are tested on 21 functions. Simulation results show that MOEA/D-IS has good performance in diversity and convergence accuracy.
{"title":"Multi-objective evolutionary algorithm based on decomposition with integration strategy","authors":"Xinwen Fang, Yuan xia Shen, Xue Feng Zhang","doi":"10.1145/3507548.3507581","DOIUrl":"https://doi.org/10.1145/3507548.3507581","url":null,"abstract":"To improve the precision in the later stage of population evolution for multi-objective evolutionary algorithm based on decomposition (MOEA/D), a MOEA/D with integration strategy (MOEA/D-IS) is proposed. The proposed algorithm adopts multiple updating strategies, including a novel first-order differential learning strategy, the individual learning strategy, and the binary and polynomial crossover mutation strategy. The penalty-based boundary intersection approach and Chebyshev approach are used to alternately evaluate individuals. The proposed algorithm and five improved MOEA algorithms are tested on 21 functions. Simulation results show that MOEA/D-IS has good performance in diversity and convergence accuracy.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"46 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128095342","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}
In the era of big data, analyzing vessels patterns using massive trajectory data has become the main method of mining activity pattern. Trajectory shape feature, as one of the important features of vessel trajectory data, can be used to identify the vessel activity patterns. But most of research only focused on the features such as standard deviation of latitude and longitude, navigation heading to the analysis of vessels trajectories. Therefore, considering the spatial-temporal feature of vessels data, we propose a method based on Sevcik fractal dimension to extract shape feature for identifying vessels activity types. Firstly, we segment the vessel trajectories to form the sub-trajectory according to the speed and temporal threshold. Secondly, we construct the feature vector of trajectory shape using the improved Sevcik fractal dimension algorithm. Then, we select the standard deviation of latitude and longitude and shape features extracted by Sevcik fractal dimension as the comparison features, and observe the performance in K-means and GMM algorithms respectively to verify the effectiveness of shape feature vectors we proposed. Finally, we select the simulation data and two real data sets for experimental analysis. The results show that the shape feature extraction algorithm can extract the shape features of trajectories, and the performance in classification algorithm is better than the standard deviation and Sevcik fractal dimension. So the method we proposed can realize the pattern recognition of vessel and abnormal trajectory analysist.
{"title":"Vessel Pattern Recognition Using Trajectory Shape Feature","authors":"Jia Li, Haiyan Liu, Xiaohui Chen, Jing Li, Junhong Xiang","doi":"10.1145/3507548.3507561","DOIUrl":"https://doi.org/10.1145/3507548.3507561","url":null,"abstract":"In the era of big data, analyzing vessels patterns using massive trajectory data has become the main method of mining activity pattern. Trajectory shape feature, as one of the important features of vessel trajectory data, can be used to identify the vessel activity patterns. But most of research only focused on the features such as standard deviation of latitude and longitude, navigation heading to the analysis of vessels trajectories. Therefore, considering the spatial-temporal feature of vessels data, we propose a method based on Sevcik fractal dimension to extract shape feature for identifying vessels activity types. Firstly, we segment the vessel trajectories to form the sub-trajectory according to the speed and temporal threshold. Secondly, we construct the feature vector of trajectory shape using the improved Sevcik fractal dimension algorithm. Then, we select the standard deviation of latitude and longitude and shape features extracted by Sevcik fractal dimension as the comparison features, and observe the performance in K-means and GMM algorithms respectively to verify the effectiveness of shape feature vectors we proposed. Finally, we select the simulation data and two real data sets for experimental analysis. The results show that the shape feature extraction algorithm can extract the shape features of trajectories, and the performance in classification algorithm is better than the standard deviation and Sevcik fractal dimension. So the method we proposed can realize the pattern recognition of vessel and abnormal trajectory analysist.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"40 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132670301","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}
Traffic sign recognition has a wide application prospect in the field of automatic driving. External factors such as illumination, Angle and occlusion will affect the recognition effect of small traffic signs. In order to solve these problems, this paper designs a multi-scale fusion convolutional neural network model (SQ-RCNN) based on feature extraction network Faster RCNN. Firstly, the multi-scale Atrous Spatial Pyramid Pooling (SASPP) module is added to the basic feature extraction network. After multi-scale cavity convolution sampling, the amount of information under each feature is not changed. In this way, the loss of resolution can be reduced and the context information of the same image can be captured. Secondly, the combination structure of two convolution layers and one pooling layer in the original VGG16 model was improved, and the concat operation was adopted to enrich the number of features by merging the number of channels, so as to realize the fusion of features at different scales and improve the accuracy of identifying small targets. In addition, a dropout layer is added to prevent overfitting. The experimental results show that: In this paper, a new network structure SQ-RCNN was used to extract features from CCTSDB data set, the mean average accuracy of traffic sign identification reached 86.96%, at the same time, effectively shorten the training time.
{"title":"Research on Traffic Sign Recognition based on Convolutional Neural Network","authors":"Wanjun Liu, Jiaxin Li, Haicheng Qu","doi":"10.1145/3507548.3507559","DOIUrl":"https://doi.org/10.1145/3507548.3507559","url":null,"abstract":"Traffic sign recognition has a wide application prospect in the field of automatic driving. External factors such as illumination, Angle and occlusion will affect the recognition effect of small traffic signs. In order to solve these problems, this paper designs a multi-scale fusion convolutional neural network model (SQ-RCNN) based on feature extraction network Faster RCNN. Firstly, the multi-scale Atrous Spatial Pyramid Pooling (SASPP) module is added to the basic feature extraction network. After multi-scale cavity convolution sampling, the amount of information under each feature is not changed. In this way, the loss of resolution can be reduced and the context information of the same image can be captured. Secondly, the combination structure of two convolution layers and one pooling layer in the original VGG16 model was improved, and the concat operation was adopted to enrich the number of features by merging the number of channels, so as to realize the fusion of features at different scales and improve the accuracy of identifying small targets. In addition, a dropout layer is added to prevent overfitting. The experimental results show that: In this paper, a new network structure SQ-RCNN was used to extract features from CCTSDB data set, the mean average accuracy of traffic sign identification reached 86.96%, at the same time, effectively shorten the training time.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"405 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133048654","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}
Andrew Kwok-Fai Lui, Y. Chan, K. Lo, Wang-To Cheng, Hang-Tak Cheung
The screening of road accident black spots is to predict accident prone locations in the road network, with the aim of preventing further accidents with remedial measures. As black spots are linked to a location, certain features of the location and its nearby branches of the network should be capable of explaining the black spots. Several open data sources now provide feature-rich road network and facilities datasets. This paper proposes a data-driven machine learning solution for black spot screening using features of road network and facilities. The accident neighborhood is a concept introduced in the paper that represents the nearby locations associated with the happening of accidents. The concept has been realized as graph embeddings of road network, which, together with a deep neural network classifier, are the two major components of the solution. An evaluation of the solution using data from a Hong Kong district indicates that recognition of both the surrounding road network structure and the local features near accident sites can yield accurate models for black spot prediction.
{"title":"Predictive Screening of Accident Black Spots based on Deep Neural Models of Road Networks and Facilities: A Case Study based on a District in Hong Kong","authors":"Andrew Kwok-Fai Lui, Y. Chan, K. Lo, Wang-To Cheng, Hang-Tak Cheung","doi":"10.1145/3507548.3507613","DOIUrl":"https://doi.org/10.1145/3507548.3507613","url":null,"abstract":"The screening of road accident black spots is to predict accident prone locations in the road network, with the aim of preventing further accidents with remedial measures. As black spots are linked to a location, certain features of the location and its nearby branches of the network should be capable of explaining the black spots. Several open data sources now provide feature-rich road network and facilities datasets. This paper proposes a data-driven machine learning solution for black spot screening using features of road network and facilities. The accident neighborhood is a concept introduced in the paper that represents the nearby locations associated with the happening of accidents. The concept has been realized as graph embeddings of road network, which, together with a deep neural network classifier, are the two major components of the solution. An evaluation of the solution using data from a Hong Kong district indicates that recognition of both the surrounding road network structure and the local features near accident sites can yield accurate models for black spot prediction.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"96 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116272457","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}
Due to the randomness and non-periodic nature of the future posture of the human body, the prediction of the posture of the human body has always been a very challenging task. In the latest research, graph convolution is proved to be an effective method to capture the dynamic relationship between the human body posture joints, which is helpful for the human body posture prediction. Moreover, graph convolution can abstract the pose of the human body to obtain a multi-scale pose set. As the level of abstraction increases, the posture movement will become more stable. Although the average prediction accuracy has improved significantly in recent years, there is still much room for exploration in the application of graph convolution in pose prediction. In this work, we propose a new multi-scale feature suppression attention map convolutional network (AZY-GCN) for end-to-end human pose prediction tasks. We use GCN to extract features from the fine-grained scale to the coarse-grained scale and then from the coarse-grained scale to the fine-grained scale. Then we combine and decode the extracted features at each scale to obtain the residual between the input and the target pose. We also performed intermediate supervision on all predicted poses so that the network can learn more representative features. In addition, we also propose a new feature suppression attention module (FISA-block), which can effectively extract relevant information from neighboring nodes while suppressing poor GCN learning noise. Our proposed method was evaluated on the public data sets of Human3.6M and CMU Mocap. After a large number of experiments, it is shown that our method has achieved relatively advanced performance.
{"title":"AZY-GCN: Multi-scale feature suppression attentional diagram convolutional network for human pose prediction","authors":"Yang Zhang, Fan Xiao Shan, Gang He","doi":"10.1145/3507548.3507565","DOIUrl":"https://doi.org/10.1145/3507548.3507565","url":null,"abstract":"Due to the randomness and non-periodic nature of the future posture of the human body, the prediction of the posture of the human body has always been a very challenging task. In the latest research, graph convolution is proved to be an effective method to capture the dynamic relationship between the human body posture joints, which is helpful for the human body posture prediction. Moreover, graph convolution can abstract the pose of the human body to obtain a multi-scale pose set. As the level of abstraction increases, the posture movement will become more stable. Although the average prediction accuracy has improved significantly in recent years, there is still much room for exploration in the application of graph convolution in pose prediction. In this work, we propose a new multi-scale feature suppression attention map convolutional network (AZY-GCN) for end-to-end human pose prediction tasks. We use GCN to extract features from the fine-grained scale to the coarse-grained scale and then from the coarse-grained scale to the fine-grained scale. Then we combine and decode the extracted features at each scale to obtain the residual between the input and the target pose. We also performed intermediate supervision on all predicted poses so that the network can learn more representative features. In addition, we also propose a new feature suppression attention module (FISA-block), which can effectively extract relevant information from neighboring nodes while suppressing poor GCN learning noise. Our proposed method was evaluated on the public data sets of Human3.6M and CMU Mocap. After a large number of experiments, it is shown that our method has achieved relatively advanced performance.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"158 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123613339","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}
Shiming Chen, Chunjing Xiao, Yanhui Han, Xianghe Du
WiFi Chanel State Information (CSI)-based activity recognition has attracted much attention in recent years. And it is extremely vital to recognize activities in time, especially for dangerous activities such as fall. In this paper, we present a real-time activity recognition system. In this system, we design a dynamic threshold-based activity segmentation method, which can address the problems of the fixed threshold and single window, and accurately detect start and end points of activities. The experiments demonstrate that our system acquires expected recognition performance.
{"title":"A Real-time Activity Recognition System based on Dynamic Adaptive Windows using WiFi Signals","authors":"Shiming Chen, Chunjing Xiao, Yanhui Han, Xianghe Du","doi":"10.1145/3507548.3507566","DOIUrl":"https://doi.org/10.1145/3507548.3507566","url":null,"abstract":"WiFi Chanel State Information (CSI)-based activity recognition has attracted much attention in recent years. And it is extremely vital to recognize activities in time, especially for dangerous activities such as fall. In this paper, we present a real-time activity recognition system. In this system, we design a dynamic threshold-based activity segmentation method, which can address the problems of the fixed threshold and single window, and accurately detect start and end points of activities. The experiments demonstrate that our system acquires expected recognition performance.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122923689","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}
Yuncheng Sang, Han Huang, Shuangqing Ma, Shouwen Cai, Zhen Shi
∗We present some experienced improvements to YOLOv5s for mobile robot platforms that occupy many system resources and are difficult to meet the requirements of the actual application. Firstly, the FPN + PAN structure is redesigned to replace this complex structure into a dilated residual module with fewer parameters and calculations. The dilated residual module is composed of dilated residual blocks with different dilation rates stacked together. Secondly, we switch some convolution modules to improved Ghost modules in the backbone. The improved Ghost modules concatenate feature maps obtained by convolution with ones generated by a linear transformation. Then, the two parts of feature maps are shuffled to boost information fusion. The model is trained on the COCO dataset. In this paper, mAP_0.5 is 56.1%, mAP_0.5:0.95 is 35.7%, and the speed is 6.1% faster than YOLOv5s. The experimental results show that the method can further increase the inference speed to ensure detection accuracy. It can well solve the task of object detection on the mobile robot platform.
{"title":"Lightweight Object Detection Method for Mobile Robot Platform","authors":"Yuncheng Sang, Han Huang, Shuangqing Ma, Shouwen Cai, Zhen Shi","doi":"10.1145/3507548.3507550","DOIUrl":"https://doi.org/10.1145/3507548.3507550","url":null,"abstract":"∗We present some experienced improvements to YOLOv5s for mobile robot platforms that occupy many system resources and are difficult to meet the requirements of the actual application. Firstly, the FPN + PAN structure is redesigned to replace this complex structure into a dilated residual module with fewer parameters and calculations. The dilated residual module is composed of dilated residual blocks with different dilation rates stacked together. Secondly, we switch some convolution modules to improved Ghost modules in the backbone. The improved Ghost modules concatenate feature maps obtained by convolution with ones generated by a linear transformation. Then, the two parts of feature maps are shuffled to boost information fusion. The model is trained on the COCO dataset. In this paper, mAP_0.5 is 56.1%, mAP_0.5:0.95 is 35.7%, and the speed is 6.1% faster than YOLOv5s. The experimental results show that the method can further increase the inference speed to ensure detection accuracy. It can well solve the task of object detection on the mobile robot platform.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127593062","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 shortcomings of traditional learning-based super-resolution (SR) reconstruction algorithms, single image super-resolution via residual dictionary learning is proposed. This method adds residual image learning to the super-resolution algorithm of beta process joint dictionary learning for coupled feature spaces. The residual dictionary pairs are learned by combining the high-resolution (HR) and low-resolution (LR) images in the external training set, which can improve the reconstruction quality and speed up the dictionary training. According to the experimental results, compared with these traditional algorithms, the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) of the proposed algorithm are significantly improved, and the visual effect is also improved.
{"title":"Single Image Super-Resolution via Residual Dictionary Learning","authors":"Yanrong Yang, Yunjie Zhang, Xiaoli Ren","doi":"10.1145/3507548.3507563","DOIUrl":"https://doi.org/10.1145/3507548.3507563","url":null,"abstract":"Aiming at the shortcomings of traditional learning-based super-resolution (SR) reconstruction algorithms, single image super-resolution via residual dictionary learning is proposed. This method adds residual image learning to the super-resolution algorithm of beta process joint dictionary learning for coupled feature spaces. The residual dictionary pairs are learned by combining the high-resolution (HR) and low-resolution (LR) images in the external training set, which can improve the reconstruction quality and speed up the dictionary training. According to the experimental results, compared with these traditional algorithms, the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) of the proposed algorithm are significantly improved, and the visual effect is also improved.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"9 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132871046","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}
With the dramatic development of information technology and rapid growth of computation performances, artificial intelligent techniques have been gradually applied in all aspects of industrial research, especially in radar signal processing. However, deep learning methods utilized in radar sea clutter are just beginning, and related researches on Doppler characteristics of sea clutter remain sparse. In this paper, artificial intelligent research on sea clutter Doppler parameters prediction is developed based on real data. Firstly, classical signal processing methods for sea clutter spectral parameters extraction are introduced. Secondly, a deep neural network model is built to predict sea clutter Doppler parameters. Finally, the raised DNN model is compared to three other classical machine learning models which are widely used in regression prediction. After comprehensive comparisons with other models in different metrics, it can be concluded that DNN model built in this paper achieves better prediction results.
{"title":"A DNN-Based Method for Sea Clutter Doppler Parameters Prediction","authors":"Xiaoyu Li, Yushi Zhang, Jinpeng Zhang","doi":"10.1145/3507548.3507595","DOIUrl":"https://doi.org/10.1145/3507548.3507595","url":null,"abstract":"With the dramatic development of information technology and rapid growth of computation performances, artificial intelligent techniques have been gradually applied in all aspects of industrial research, especially in radar signal processing. However, deep learning methods utilized in radar sea clutter are just beginning, and related researches on Doppler characteristics of sea clutter remain sparse. In this paper, artificial intelligent research on sea clutter Doppler parameters prediction is developed based on real data. Firstly, classical signal processing methods for sea clutter spectral parameters extraction are introduced. Secondly, a deep neural network model is built to predict sea clutter Doppler parameters. Finally, the raised DNN model is compared to three other classical machine learning models which are widely used in regression prediction. After comprehensive comparisons with other models in different metrics, it can be concluded that DNN model built in this paper achieves better prediction results.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"415 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133382712","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}
In the deep texture image retrieval, to address the problem that the retrieval performance is affected by the lack of sufficiently large texture image dataset used for the effective training of deep neural network, a deep learning based texture dataset construction and texture image retrieval method is proposed in this paper. First, a large-scale texture image dataset containing rich texture information is constructed based on the DTD texture image dataset, and used as the source dataset for pre-training deep neural networks. To effectively characterize the information of the source texture dataset, a revised version of the VGG16 model, called ReV-VGG16, is adaptively designed. Then, the pre-trained ReV-VGG16 model is combined with the target texture image datasets for the transfer learning, and the probability values of the output from the classification layer of the model are used for the computation of the similarity measurement to achieve the retrieval of the target texture image dataset. Finally, the retrieval experiments are conducted on four typical texture image datasets, namely, VisTex, Brodatz, STex and ALOT. The experimental results show that our method outperforms the existing state-of-the-art texture image retrieval approaches in terms of the retrieval performance.
{"title":"Texture Dataset Construction and Texture Image Retrieval based on Deep Learning","authors":"Zhisheng Zhang, Huaijing Qu, Hengbin Wang, Jia Xu, Jiwei Wang, Yanan Wei","doi":"10.1145/3507548.3507564","DOIUrl":"https://doi.org/10.1145/3507548.3507564","url":null,"abstract":"In the deep texture image retrieval, to address the problem that the retrieval performance is affected by the lack of sufficiently large texture image dataset used for the effective training of deep neural network, a deep learning based texture dataset construction and texture image retrieval method is proposed in this paper. First, a large-scale texture image dataset containing rich texture information is constructed based on the DTD texture image dataset, and used as the source dataset for pre-training deep neural networks. To effectively characterize the information of the source texture dataset, a revised version of the VGG16 model, called ReV-VGG16, is adaptively designed. Then, the pre-trained ReV-VGG16 model is combined with the target texture image datasets for the transfer learning, and the probability values of the output from the classification layer of the model are used for the computation of the similarity measurement to achieve the retrieval of the target texture image dataset. Finally, the retrieval experiments are conducted on four typical texture image datasets, namely, VisTex, Brodatz, STex and ALOT. The experimental results show that our method outperforms the existing state-of-the-art texture image retrieval approaches in terms of the retrieval performance.","PeriodicalId":414908,"journal":{"name":"Proceedings of the 2021 5th International Conference on Computer Science and Artificial Intelligence","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130991618","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：