Pub Date : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989603
Dakota Joiner, Mathias Clement, Shek Tom Chan, Keegan Pereira, Albert Wong, Y. Khmelevsky, Joe Mahony, Michael Ferri
This case study shows the performance issues and solutions for a data warehouse (DW) performing well to serve industrial partners in improving customer data retrieval performance. An online transaction processing (OLTP) relational database and a DW were deployed in PostgreSQL and tested against each other. Several test cases were carried out with the DW, including indexing and creating pre-aggregated tables, all guided by in-depth analysis of EXPLAIN plans. Queries and DW design were continually improved throughout testing to ensure that the OLTP and DW were compared equally. Seven queries (requested by the industrial client) were used to thoroughly test different performance aspects concerning client feedback and the complexity of requests for all areas the DW might cover. On average, the data warehouse showed a one to three magnitudes increase in query execution performance, with the highest calibre results coming in at 2,493 times faster than the OLTP. All test cases showed an increase in performance over the OLTP. Additionally, the data contained in the DWtook up 24% less storage space than the OLTP. The results here indicate a promising direction to take business analytics with data warehousing, as customers will experience significant cost savings and a reduction in time to receive desired results from their data storage platforms in the cloud. The work in this case study is a continuation of previous work in a much larger project concerning integrating database technologies with machine learning to improve natural language processing solutions as a cost-saving measure for utilities consumers.
{"title":"DW vs OLTP Performance Optimization in the Cloud on PostgreSQL (A Case Study)","authors":"Dakota Joiner, Mathias Clement, Shek Tom Chan, Keegan Pereira, Albert Wong, Y. Khmelevsky, Joe Mahony, Michael Ferri","doi":"10.1109/RASSE54974.2022.9989603","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989603","url":null,"abstract":"This case study shows the performance issues and solutions for a data warehouse (DW) performing well to serve industrial partners in improving customer data retrieval performance. An online transaction processing (OLTP) relational database and a DW were deployed in PostgreSQL and tested against each other. Several test cases were carried out with the DW, including indexing and creating pre-aggregated tables, all guided by in-depth analysis of EXPLAIN plans. Queries and DW design were continually improved throughout testing to ensure that the OLTP and DW were compared equally. Seven queries (requested by the industrial client) were used to thoroughly test different performance aspects concerning client feedback and the complexity of requests for all areas the DW might cover. On average, the data warehouse showed a one to three magnitudes increase in query execution performance, with the highest calibre results coming in at 2,493 times faster than the OLTP. All test cases showed an increase in performance over the OLTP. Additionally, the data contained in the DWtook up 24% less storage space than the OLTP. The results here indicate a promising direction to take business analytics with data warehousing, as customers will experience significant cost savings and a reduction in time to receive desired results from their data storage platforms in the cloud. The work in this case study is a continuation of previous work in a much larger project concerning integrating database technologies with machine learning to improve natural language processing solutions as a cost-saving measure for utilities consumers.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129177426","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989570
Haohsuan Tseng, Chih-Hung Kuo, Yiting Chen, Sinhong Lee
In this paper, we propose an edge-guided video super-resolution (EGVSR) network that utilizes the edge information of the image to effectively recover high-frequency details for high-resolution frames. The reconstruction process consists of two stages. In the first stage, the Coarse Frame Reconstruction Network (CFRN) generates coarse SR frames. In addition, we propose the Edge-Prediction Network (EPN) to capture the edge details that help to supplement the missing high-frequency information. Unlike some prior SR works that tend to increase the depth of networks or use attention mechanisms to reconstruct large-size objects but ignore small-size objects, we propose the Attention Fusion Residual Block (AFRB) to process objects of different sizes. The AFRB, an enhanced version of the conventional residual block, performs fusion through a multi-scale channel attention mechanism and serves as the basic operation unit in the CFRN and the EPN. Then, in the second stage, we propose the Frame Refinement Network (FRN), which contains multiple convolution layers. Through the FRN, we fuse and refine the coarse SR frames and edge information learned from the first stage. Compared with the state-of-the-art methods, our SR model improves approximately 0.5% in PSNR and 1.8% in SSIM evaluation on the benchmark VID4 dataset when the number of parameters is reduced by 54%.
{"title":"Edge-Guided Video Super-Resolution Network","authors":"Haohsuan Tseng, Chih-Hung Kuo, Yiting Chen, Sinhong Lee","doi":"10.1109/RASSE54974.2022.9989570","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989570","url":null,"abstract":"In this paper, we propose an edge-guided video super-resolution (EGVSR) network that utilizes the edge information of the image to effectively recover high-frequency details for high-resolution frames. The reconstruction process consists of two stages. In the first stage, the Coarse Frame Reconstruction Network (CFRN) generates coarse SR frames. In addition, we propose the Edge-Prediction Network (EPN) to capture the edge details that help to supplement the missing high-frequency information. Unlike some prior SR works that tend to increase the depth of networks or use attention mechanisms to reconstruct large-size objects but ignore small-size objects, we propose the Attention Fusion Residual Block (AFRB) to process objects of different sizes. The AFRB, an enhanced version of the conventional residual block, performs fusion through a multi-scale channel attention mechanism and serves as the basic operation unit in the CFRN and the EPN. Then, in the second stage, we propose the Frame Refinement Network (FRN), which contains multiple convolution layers. Through the FRN, we fuse and refine the coarse SR frames and edge information learned from the first stage. Compared with the state-of-the-art methods, our SR model improves approximately 0.5% in PSNR and 1.8% in SSIM evaluation on the benchmark VID4 dataset when the number of parameters is reduced by 54%.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"38 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131559765","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989723
I-hsiang Lai, Wei-Liang Lin
Modern robots need to interact with human and move around human environment, in places such as museums, restaurants, or supermarkets. Therefore, robots should have social navigation capability. This article uses object detection to detect pedestrians, fuses object detection result with lidar information to obtain the state of the pedestrian, and then changes the navigation path according to the calculated pedestrian state. When there are people face-to-face and talking to each other, the autonomous mover bypasses instead of passing through them. When pedestrian in front of the autonomous mover is crossing the autonomous mover from left to right, the autonomous mover turns left to pass the other side instead of going straight and blocking the pedestrian. Therefore, the autonomous mover can navigate without disturbing pedestrians and respect social distance.Our approach uses a single RGB camera and a one-line lidar to detect pedestrian and accomplish the two specific goals in the real world. We fuse lidar information and object detection result to obtain the position and face orientation of the pedestrian. We add a customized social layer to the cost map of an existing navigation system, and thus, change the original shortest path algorithm. The face-to-face and crossing scenarios are verified in the hall of a university department building.
{"title":"Autonomous Mover with Social Distance Respect","authors":"I-hsiang Lai, Wei-Liang Lin","doi":"10.1109/RASSE54974.2022.9989723","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989723","url":null,"abstract":"Modern robots need to interact with human and move around human environment, in places such as museums, restaurants, or supermarkets. Therefore, robots should have social navigation capability. This article uses object detection to detect pedestrians, fuses object detection result with lidar information to obtain the state of the pedestrian, and then changes the navigation path according to the calculated pedestrian state. When there are people face-to-face and talking to each other, the autonomous mover bypasses instead of passing through them. When pedestrian in front of the autonomous mover is crossing the autonomous mover from left to right, the autonomous mover turns left to pass the other side instead of going straight and blocking the pedestrian. Therefore, the autonomous mover can navigate without disturbing pedestrians and respect social distance.Our approach uses a single RGB camera and a one-line lidar to detect pedestrian and accomplish the two specific goals in the real world. We fuse lidar information and object detection result to obtain the position and face orientation of the pedestrian. We add a customized social layer to the cost map of an existing navigation system, and thus, change the original shortest path algorithm. The face-to-face and crossing scenarios are verified in the hall of a university department building.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"274 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124261705","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989881
Yu-Wen Wang, Gwo Giun Chris Lee, Yu-Hsuan Chen, Shih-Yu Chen, Tai-Ping Wang
This paper implements an application specific design for calculating the two-dimensional convolution with given Gabor filters onto a Field Programmable Gate Array (FPGA). Nowadays, Convolutional Neural Network (CNN) is a widely used algorithm in the field of computer vision. However, the amount of computation it requires is immense, and therefore special algorithms and hardware are necessary to accelerate the process. We introduce the Eigen-transformation approach, which transforms the 16 Gabor filters into another 16 filters with increased symmetry. This reduces the number of operations, as well as allows us to pre-add the input pixels corresponding to the position of the repeated coefficients. Previous works from our lab analyze the symmetry properties of 7×7 Gabor filters and build the dataflow model of Gabor filter based convolution and use software to implement it. In this paper, we analyze the four models of processing units for the transformed filter bank proposed by the previous work in our lab and use the Xilinx XUPV5-LX110T Evaluation Platform for prototyping. The proposed four models each have unique advantages that make them suitable for different applications. In the experiment, we use a 224×224 image as input and the bit-width of data is 32. Finally, we use the Xilinx Chipscope as an integrated logic analyzer for verification.
{"title":"Implementation of Gabor Filter Based Convolution for Deep Learning on FPGA","authors":"Yu-Wen Wang, Gwo Giun Chris Lee, Yu-Hsuan Chen, Shih-Yu Chen, Tai-Ping Wang","doi":"10.1109/RASSE54974.2022.9989881","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989881","url":null,"abstract":"This paper implements an application specific design for calculating the two-dimensional convolution with given Gabor filters onto a Field Programmable Gate Array (FPGA). Nowadays, Convolutional Neural Network (CNN) is a widely used algorithm in the field of computer vision. However, the amount of computation it requires is immense, and therefore special algorithms and hardware are necessary to accelerate the process. We introduce the Eigen-transformation approach, which transforms the 16 Gabor filters into another 16 filters with increased symmetry. This reduces the number of operations, as well as allows us to pre-add the input pixels corresponding to the position of the repeated coefficients. Previous works from our lab analyze the symmetry properties of 7×7 Gabor filters and build the dataflow model of Gabor filter based convolution and use software to implement it. In this paper, we analyze the four models of processing units for the transformed filter bank proposed by the previous work in our lab and use the Xilinx XUPV5-LX110T Evaluation Platform for prototyping. The proposed four models each have unique advantages that make them suitable for different applications. In the experiment, we use a 224×224 image as input and the bit-width of data is 32. Finally, we use the Xilinx Chipscope as an integrated logic analyzer for verification.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129415917","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989732
Sing-Yu Pan, Shuenn-Yuh Lee, Yi-Wen Hung, Chou-Ching K. Lin, G. Shieh
This paper has proposed an epilepsy detection algorithm to identify the seizure attack. The algorithm includes a simplified signal preprocessing process and an 8 layers Convolution Neural Network (CNN). This paper has also proposed an architecture, including a CNN accelerator and a 2-stage reduced instruction set computer-V (RISC-V) CPU, to implement the detection algorithm in real-time. The accelerator is implemented in SystemVerilog and validated on the Xilinx PYNQ-Z2. The implementation consumes 3411 LUTs, 2262 flip-flops, 84 KB block random access memory (BRAM), and only 6 DSPs. The total power consumption is 0.118 W in 10-MHz operation frequency. The detection algorithm provides 99.16% accuracy on fixed-point operations with detection latency of 0.137 ms/class. Moreover, the CNN accelerator has the programable ability, so the accelerator can execute different CNN models to fit various wearable applications for different biomedical acquisition systems.
本文提出了一种癫痫检测算法来识别癫痫发作。该算法包括一个简化的信号预处理过程和一个8层卷积神经网络(CNN)。本文还提出了一种包括CNN加速器和二级精简指令集计算机- v (RISC-V) CPU的架构，以实现实时检测算法。该加速器在SystemVerilog中实现，并在Xilinx PYNQ-Z2上进行了验证。该实现消耗3411个lut、2262个触发器、84 KB块随机存取存储器(BRAM)和仅6个dsp。在10mhz工作频率下，总功耗为0.118 W。该算法对定点运算的检测准确率为99.16%，检测延迟为0.137 ms/class。此外，CNN加速器具有可编程能力，因此加速器可以执行不同的CNN模型，以适应不同生物医学采集系统的各种可穿戴应用。
{"title":"A Programmable CNN Accelerator with RISC-V Core in Real-Time Wearable Application","authors":"Sing-Yu Pan, Shuenn-Yuh Lee, Yi-Wen Hung, Chou-Ching K. Lin, G. Shieh","doi":"10.1109/RASSE54974.2022.9989732","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989732","url":null,"abstract":"This paper has proposed an epilepsy detection algorithm to identify the seizure attack. The algorithm includes a simplified signal preprocessing process and an 8 layers Convolution Neural Network (CNN). This paper has also proposed an architecture, including a CNN accelerator and a 2-stage reduced instruction set computer-V (RISC-V) CPU, to implement the detection algorithm in real-time. The accelerator is implemented in SystemVerilog and validated on the Xilinx PYNQ-Z2. The implementation consumes 3411 LUTs, 2262 flip-flops, 84 KB block random access memory (BRAM), and only 6 DSPs. The total power consumption is 0.118 W in 10-MHz operation frequency. The detection algorithm provides 99.16% accuracy on fixed-point operations with detection latency of 0.137 ms/class. Moreover, the CNN accelerator has the programable ability, so the accelerator can execute different CNN models to fit various wearable applications for different biomedical acquisition systems.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129799081","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989592
Dakota Joiner, Amy Vezeau, Albert Wong, Gaétan Hains, Y. Khmelevsky
Stock price prediction with machine learning is an oft-studied area where numerous unsolved problems still abound owing to the high complexity and volatility that technical-factors and sentiment-analysis models are trying to capture. Nearly all areas of machine learning (ML) have been tested as solutions to generate a truly accurate predictive model. The accuracy of most models hovers around 50%, highlighting the need for further increases in precision, data handling, forecasting, and ultimately prediction. This literature review aggregates and concludes the current state of the art (from 2018 onward) with specifically selected criteria to guide further research into algorithmic trading. The review targets academic papers on ML or deep learning (DL) with algorithmic trading or data sets used for algorithmic trading with minute to daily time scales. Systems that integrate and test sentiment and technical analysis are considered the best candidates for an eventual generalized trading algorithm that can be applied to any stock, future, or traded commodity. However, much work remains to be done in applying natural language processing and the choice of text sources to find the most effective mixture of sentiment and technical analysis. The best models being useless on themselves, we also search for publications about data warehousing systems aggregating financial factors impacting stock prices. A brief review in this area is included in this regard.
{"title":"Algorithmic Trading and Short-term Forecast for Financial Time Series with Machine Learning Models; State of the Art and Perspectives","authors":"Dakota Joiner, Amy Vezeau, Albert Wong, Gaétan Hains, Y. Khmelevsky","doi":"10.1109/RASSE54974.2022.9989592","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989592","url":null,"abstract":"Stock price prediction with machine learning is an oft-studied area where numerous unsolved problems still abound owing to the high complexity and volatility that technical-factors and sentiment-analysis models are trying to capture. Nearly all areas of machine learning (ML) have been tested as solutions to generate a truly accurate predictive model. The accuracy of most models hovers around 50%, highlighting the need for further increases in precision, data handling, forecasting, and ultimately prediction. This literature review aggregates and concludes the current state of the art (from 2018 onward) with specifically selected criteria to guide further research into algorithmic trading. The review targets academic papers on ML or deep learning (DL) with algorithmic trading or data sets used for algorithmic trading with minute to daily time scales. Systems that integrate and test sentiment and technical analysis are considered the best candidates for an eventual generalized trading algorithm that can be applied to any stock, future, or traded commodity. However, much work remains to be done in applying natural language processing and the choice of text sources to find the most effective mixture of sentiment and technical analysis. The best models being useless on themselves, we also search for publications about data warehousing systems aggregating financial factors impacting stock prices. A brief review in this area is included in this regard.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"40 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129859841","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989693
Holly A. H. Handley, A. Marnewick
The Accreditation Board for Engineering and Technology (ABET) has included the principles of diversity, equity, and inclusion (DEI) into the General Criteria for Accrediting Engineering Programs. The intent is for these professional competencies to be taught in tandem with the technical skills provided by the engineering curriculum, however there are few guidelines on how to do this. In this paper, the authors propose a model to help instructors incorporate DEI principles into existing engineering coursework; the approach is based on a competency building model previously developed to integrate international competencies into systems engineering courses by identifying opportunities in the student learning cycle. This paper adapts that model by identifying both the appropriate competencies to develop and the classroom context factors that support DEI. The competencies, cognitive style awareness and teamwork, are fostered using constructs from a learning model, while the classroom context factors include the technical content, interactions with other students, and the teaching environment. An example is provided that illustrates the use of the model in a systems engineering curriculum to improve an existing course module to better adhere to DEI principles. The DEI competencies identified by this model augment those advocated by the systems engineering community, i.e., the necessary characteristics to be successful as professionals in the systems engineering field. The integrated DEI model can be used to improve existing curriculum to meet these needs.
{"title":"A Diversity, Equity, and Inclusion Model for Engineering Curriculums","authors":"Holly A. H. Handley, A. Marnewick","doi":"10.1109/RASSE54974.2022.9989693","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989693","url":null,"abstract":"The Accreditation Board for Engineering and Technology (ABET) has included the principles of diversity, equity, and inclusion (DEI) into the General Criteria for Accrediting Engineering Programs. The intent is for these professional competencies to be taught in tandem with the technical skills provided by the engineering curriculum, however there are few guidelines on how to do this. In this paper, the authors propose a model to help instructors incorporate DEI principles into existing engineering coursework; the approach is based on a competency building model previously developed to integrate international competencies into systems engineering courses by identifying opportunities in the student learning cycle. This paper adapts that model by identifying both the appropriate competencies to develop and the classroom context factors that support DEI. The competencies, cognitive style awareness and teamwork, are fostered using constructs from a learning model, while the classroom context factors include the technical content, interactions with other students, and the teaching environment. An example is provided that illustrates the use of the model in a systems engineering curriculum to improve an existing course module to better adhere to DEI principles. The DEI competencies identified by this model augment those advocated by the systems engineering community, i.e., the necessary characteristics to be successful as professionals in the systems engineering field. The integrated DEI model can be used to improve existing curriculum to meet these needs.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122220448","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989741
Chin-Chieh Chang, Wei-Liang Ou, Hua-Luen Chen, Chih-Peng Fan
In this study, a YOLO-based deep-learning gaze estimation technology is developed for the application of non-contact smart advertising displays. By integrating the appearance and geometric-features technologies, the output coordinates of facial features inferred by YOLOv3-tiny based models can provide the training data for gaze estimation without the calibration process. In experiments, the input size of YOLOv3-tiny based models is arranged by 608x608 pixels, and the used models have good location performance to detect the facial directions and two facial features. By the YOLOv3-tiny based model with the cross-person test, the proposed method performs the averaged gaze estimation accuracies of nine, six, and four-block modes are 66.38%, 80.87%, 88.34%, respectively with no calibration process.
{"title":"YOLO-Based Deep-Learning Gaze Estimation Technology by Combining Geometric Feature and Appearance Based Technologies for Smart Advertising Displays","authors":"Chin-Chieh Chang, Wei-Liang Ou, Hua-Luen Chen, Chih-Peng Fan","doi":"10.1109/RASSE54974.2022.9989741","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989741","url":null,"abstract":"In this study, a YOLO-based deep-learning gaze estimation technology is developed for the application of non-contact smart advertising displays. By integrating the appearance and geometric-features technologies, the output coordinates of facial features inferred by YOLOv3-tiny based models can provide the training data for gaze estimation without the calibration process. In experiments, the input size of YOLOv3-tiny based models is arranged by 608x608 pixels, and the used models have good location performance to detect the facial directions and two facial features. By the YOLOv3-tiny based model with the cross-person test, the proposed method performs the averaged gaze estimation accuracies of nine, six, and four-block modes are 66.38%, 80.87%, 88.34%, respectively with no calibration process.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"403 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114938807","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 : 2022-11-07DOI: 10.1109/RASSE54974.2022.9989667
Hsi-Ling Chen, J. Yang, Song-An Mao
CNN models are becoming more and more mature, many of them adopt deeper structures to better accomplish the task objectives, such that the increased computational and storage burdens are unfavorable for the implementation in edge devices. In this paper, we propose an approach to optimize the filter structure by starting from the convolutional filter and finding their minimum structure. The reductions of the filters for the minimum structure in terms of space and channels, the number of model parameters and the computational complexity are effectively reduced. Since the current channel pruning method prunes the same channel for each convolutional layer, which easily leads to a trade-off between the pruning rate and accuracy loss. Instead we propose a new channel pruning approach to find the most suitable required channels for each filter to provide a more detailed pruning method. Experiments conducted on the classification CNN models, such as VGG16 and ResNet56, show that the proposed method can successfully reduce the computations of the models without losing much model accuracy effectively. The proposed method performs well in compressing the model and reducing the number of parameters required by the models for real applications.
{"title":"Convolutional Layers Acceleration By Exploring Optimal Filter Structures","authors":"Hsi-Ling Chen, J. Yang, Song-An Mao","doi":"10.1109/RASSE54974.2022.9989667","DOIUrl":"https://doi.org/10.1109/RASSE54974.2022.9989667","url":null,"abstract":"CNN models are becoming more and more mature, many of them adopt deeper structures to better accomplish the task objectives, such that the increased computational and storage burdens are unfavorable for the implementation in edge devices. In this paper, we propose an approach to optimize the filter structure by starting from the convolutional filter and finding their minimum structure. The reductions of the filters for the minimum structure in terms of space and channels, the number of model parameters and the computational complexity are effectively reduced. Since the current channel pruning method prunes the same channel for each convolutional layer, which easily leads to a trade-off between the pruning rate and accuracy loss. Instead we propose a new channel pruning approach to find the most suitable required channels for each filter to provide a more detailed pruning method. Experiments conducted on the classification CNN models, such as VGG16 and ResNet56, show that the proposed method can successfully reduce the computations of the models without losing much model accuracy effectively. The proposed method performs well in compressing the model and reducing the number of parameters required by the models for real applications.","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"10 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133756093","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 : 2022-11-07DOI: 10.1109/rasse54974.2022.9989913
{"title":"RASSE 2022 Cover Page","authors":"","doi":"10.1109/rasse54974.2022.9989913","DOIUrl":"https://doi.org/10.1109/rasse54974.2022.9989913","url":null,"abstract":"","PeriodicalId":382440,"journal":{"name":"2022 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE)","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131438680","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：