.

  8:50-9:00 (Bangkok Time) Opening Ceremony Chaired by Prof Sansanee Auephanwiriyakul, Winter School/DL Chair

1, Introduction of the DLP/Winter school, Speakers and co-organizer   by Prof. S Auephanwiriyakul

2, Welcome Speech and announce the opening of the school, by APSIPA VP and Winter School/DL Chair, Prof. Mingyi He

9:00-10:00  Chaired by Prof. Sansanee Auephanwiriyakul

Lecture 1

Speaker

Prof. Ye Wang

Abstract

Speech is an essential means for human interactions and music is more than just a source of entertainment. In this talk, I will introduce our recent work to demonstrate how speech and music information processing techniques can enable novel applications in promoting Human Health and Potential (HHP). Our first work MusicRx-C focuses on generating user-preferred music for gait rehabilitation of patients with Parkinson’s disease. Our second work AI-lyricist is a machine learning based system to generate lyrics for language learning. Our third work extends AI-Lyricist into a Karaoke application SLIONS, specially designed for promoting bilingual education in Singapore. We have attempted to transform insights gained from cognitive psychology and neuroscience into real-life applications that can be delivered widely and effectively. The key idea behind the project is to combine teachers’ human intelligence with big data, and to convert them into machine intelligence which will facilitate joyful and effective learning. The SLIONS’ activity and feedback components incorporate features designed to support self-regulated learning as well as features informed by variation theory and multimedia learning theory. I will also discuss the joys and pains working in such a multidisciplinary environment, and will end with some possible research directions.

10:00-10:15 Break

10:15-11:15   Chaired by Prof. Mingyi He

Lecture 2

Trainable Subspaces for Tensor Completion

Prof. Yipeng Liu

Abstract

With the help of auxiliary data, tensor completion can better recover a multidimensional array from limited observations. Existing methods are either sensitive to a given rank or lack of physical interpretations of subspace information. In this work, we propose to separately exploit shared subspaces for tensor completion. Specifically, dictionary learning takes the subspace from auxiliary data in the first step. Then a low rank optimization model for tensor completion is provided to incorporate the trained subspace by assuming that the recovered tensor is composed of two low rank components where one shares the subspace information with auxiliary data and the other is outside the shared space. Based on this optimization model, we make a quantitative analysis to illustrate the effect of subspace information on sample complexity, and provide theoretical insights into the usefulness of subspace information. Finally, experiments on simulated data are conducted to validate the theoretical analysis on the impact of subspace information，and experiments in two real-world image recovery applications show the proposed method outperforms state-of-the-art ones in terms of prediction accuracy and computational complexity.

Yipeng Liu is an associate professor with University of Electronic Science and Technology of China (UESTC), Chengdu, China. His research interest is tensor for data processing. He has published over 70 papers, co-authored two books “Tensor Computation for Data Analysis” by Springer and “Tensor Regression” by NOW Publishers, and edited one book “Tensors for Data Processing” by Elsevier. He has been served as an associate editor for IEEE Signal Processing Letters. He has given tutorials for several international conferences, such as IJCAI 2022, ICME 2022, MLSP 2022, VCIP 2021, ICIP 2020, SSCI 2020, ISCAS 2019, SiPS 2019, and APSIPA ASC 2019. He is a APSIPA Distinguished Lecturer 2022-2023.

 11:15-12:15   Chaired by Prof. Navadon Khunlertgit

Lecture 3

Automated Landslide-Risk Prediction based Extreme Class-Imbalance Dataset using Two-Stage-Transfer Learning

Speaker

Prof. Naruephorn Tengtrairat

Abstract

Global warming has caused rapid climate change and has caused catastrophic natural disasters in the past few years. Landslides occur more frequently and have a significant crucial immense harm on life and property. Based on the rapid development of machine learning principles, this has therefore led to a widely proposed landslide forecasting method with improved forecasting performance.

When the relatively small or limited number of landslide data are available that will lead to an extreme scenario of the landslide prediction problem. To overcome the limitations of the imbalance landslide dataset and landslide-factor limitation, the landslide prediction method will be introduced to capture the significant features from the training model process. The state-of-the-art machine learning approaches for building landslide prediction will be discussed and then demonstrated the landslide prediction-based web GIS applications. Finally, vital factors for the development of landslide forecasting and future development directions will be presented.

Biography

Dr. Naruephorn Tengtrairat received the B.Eng. degree in computer engineering from the Chiang Mai University, Chiang Mai, Thailand, then M.Sc. degree in management information system from the Chulalongkorn University, and Ph.D. from Newcastle University, U.K. She has been a lecturer and Head of the department of Software Engineering at Payap University, Thailand. Her research interests include statistical single-channel blind source separation, speech and audio signal processing, speech enhancement, noise cancelling, machine learning, and machine vision.

Dr. Naruephorn Tengtrairat is currently serving as the technical committee of AIVRV 2022, ICSPS 2022, AASIP 2022, Eusipco2014, and including as the organizing committees of CAIBDA 2021. She serves as a reviewer to several international journals. She was a co-founder of a start-up company ‘ReadRing’ in the section of smart microcontrollers and AI software on optical character recognition translation to braille representation.

 12:15-13:30  Lunch Time

13:30-14:30  Chaired by Prof. Natthanan Promsuk

Lecture 4

Fine-grained Human Action Recognition

Speaker

Prof. Zhiyong Wang

Abstract

Human action recognition from videos has been well researched for decades due to numerous applications, such as surveillance, sports, and healthcare. Due to the groundbreaking success of deep learning, a new challenge has recently emerged in the field, namely, fine-grained human action recognition which aims to recognize subtle actions often only recognizable by well-trained experts in realistic scenarios. In this talk, we will first briefly review the field of human action recognition to highlight the challenges, and then present recent progress on fine-grained human action with various deep learning techniques.

Biography

Zhiyong Wang is an Associate Professor and Director of the Multimedia Computing Laboratory at the School of Computer Science, The University of Sydney, Australia. He received his B. Eng. and M. Eng. Degrees in electronic engineering from South China University of Technology, Guangzhou, China, and his Ph.D. degree from Hong Kong Polytechnic University, Hong Kong. His research interests focus on multimedia computing and its applications in agriculture, earth observation, health, and medicine, including multimedia information processing, analysis, retrieval, summarization, and recognition (e.g., human action recognition and affective analysis), and multimedia data mining, multimedia content creation, human-centered multimedia computing, remote sensing, and pattern recognition. He is an Associate Editor of ACM Computing Surveys and Neurocomputing, and was the President of Australia Pattern Recognition Society (APRS), and APSIPA Distinguished Lecturer 2021-2022.

14:30-14:45 Break

14:45-15:45 Chaired by Prof. Kampol Woradit

Lecture 5

High Framerate Video Generation from Rolling Shutter Cameras and Event Cameras

Speaker

Prof. Yuchao Dai

Abstract

Dynamic is ubiquitous in the 3D world, ranging from the driving vehicles, the running water, to the dancing people. Different dynamic scenes recordings require different observation frame rates. The current frame rate of mobile phone imaging systems has been continuously improved. However, there is still a gap between the requirements for high frame rate and high resolution recording. In this lecture, I will focus on our research group’s recent work on high frame rate video generation, mainly covering the topics of high frame rate video generation based on event camera and low frame rate video, and high frame rate video generation based on rolling shutter cameras. I will conclude the lecture with discussions on open issues and recent trends in this field.

Biography

Yuchao Dai is currently a Professor with School of Electronics and Information at the Northwestern Polytechnical University (NPU), Xi’an, China. He received the B.E. degree, M.E degree and Ph.D. degree all in Signal and Information Processing from NPU, in 2005, 2008 and 2012, respectively. He was an ARC DECRA Fellow with the Research School of Engineering at the Australian National University, Canberra, Australia. His research interests include structure from motion, multi-view geometry, low-level computer vision, deep learning, compressive sensing and optimization. He has published more than 100 papers in TPAMI, IJCV, CVPR, ICCV, ECCV, NeurIPS and etc. He won the Best Paper Award at IEEE CVPR 2012, the Best Paper Award Nominee at IEEE CVPR 2020, the DSTO Best Fundamental Contribution to Image Processing Paper Prize at DICTA 2014, the Best Algorithm Prize in NRSFM Challenge at CVPR 2017, the Best Student Paper Prize at DICTA 2017 and the Best Deep/Machine Learning Paper Prize at APSIPA ASC 2017. He served as Area Chair for IEEE CVPR, ICCV, ACM MM and etc. He serves as the Publicity Chair for ACCV 2022. He is a APSIPA Distinguished Lecturer 2022-2023.

15:45-16:45  Chaired by Prof. Thanatip Chankong

Lecture 6

Applications of localised AI technology in Thailand

Speaker

Dr. Thanawat Thiasiriphet

Abstract

The Artificial Intelligence technology has been continuously grown and rapidly advanced every single day. For globally-scaled research works, we have seen some of the ‘new’ technologies became ‘old’ within a few months. However, horizontal growth (localization) in AI technology cannot catch up with such vertical growth. Virtual agents as well as AI agents developed by world’s top companies are good but fairly limited for deploying in Thailand. IBotnoi Co. LTD. is one of the leading AI companies in Thailand which aims to fill these gaps. Being a local company, there are both chance and expertise in collecting specific data in several domains, and therefore more specialized and localized AI services can be available. This talk give a brief review about Botnoi’s service as well as the theories behind.

Biography

Thanawat Thiasiriphet was born in Chiang mai, Thailand in 1982. He is currently working as Managing Director at iBotnoi Co. LTD. He received his Ph. D. degree in electrical engineering from University on Ulm, Germany in 2014. Since then he has been working in both academia and industry in Germany and Thailand.

16:45-17:45 Chaired by Prof. Sansanee Auephanwiriyakul

Lecture 7

Learning based Methods for Hyperspectral Big Data Processing

Speaker

Prof. Mingyi HE

Abstract

Hyperspectral imaging is a new imaging technology formed by the combination of digital imaging and high-resolution spectral separation techniques. It can obtain two-dimensional images on many narrow spectral channels that are close to continuous, so as to collect the bigdata of detail spectral and spatial information of the objects and their environments. With hyperspectral bigdata, it is possible to recognize the 2D shapes of observed objects and their materials and even break through the physical limitation of 2D spatial resolution to analyze the mixed objects, for example the analysis of the mixed powder medicine. Therefore, they have broad application values in various quantitative remote sensing, material identification, medical diagnosis, quality monitoring, and camouflage detection, etc. It is still listed internationally as one of the fields with breakthrough and subversive development. Due to the redundancy, coupling and correlation of hyperspectral images in space, spectral and time, there are many problems such as dimension disaster, the phenomena of different objects with same spectrum and same objects with different spectra, thus hyperspectral data processing is a great challenge. In the past decade, hyperspectral image data processing has been highly valued and made important progress in the world with learning based methods especially with deep learning. This talk will introduce the learning based methods for hyperspectral remote sensing data processing with applications in classification, fusion, and mixture analysis. These methods include the shallow feature learning methods (such as feature mining with multilayer feedforward neural network-MLFNN; double parallel feedforward NN-DPFNN, an MLFNN with a skip connection; etc), the end-to-end deep learning methods (such as DCNN and skip deep neural networks:  U-Net, ResNet, Multi-scale 3D DCNN, etc), and active learning based methods. The future trends will also be discussed.

Biography

Mingyi He, professor, School of Electronic and Information, Northwestern Polytechnical University, director of Shaanxi International Joint Research Center for Information Acquisition and Processing (IAP), and funding director of key lab of IAP. His research has been concerned with the cross fields of signal and information processing, machine vision and image processing, neural network artificial intelligence, and hyperspectral remote sensing data processing. He has published “Neural Networks and Signal Processing”(NPU Press, 1998), Digital Image Processing(Science Press, 2008), more than 300 journal and conference papers in IEEE TGRS, TPAMI, IJCV, TGRS, PR, ICIP, IGARSS, etc. He has won over 20 international/national academic awards such as CVPR 2012 best paper award, 11 scientific prizes/3 teaching prizes from China and the governmental recognition of outstanding contribution to higher education and scientific research by the state council of China. He is a vice president of APSIPA, a Fellow of Chinese Institute of Electronics, and vice president of Shaanxi Institute of Electronics. He has served as the editorial board member or guest editor for IEEE TGRS, Jstars, Remote Sensing, Journal of Signal Processing, Chinese Journal of Image and Graphics, etc.

17:45-18:00    Closing Remark