Project
Computer vision
Machine vision for chess playing robot
A robotic vision module is developed for providing the information of chess board and pieces by using a camera
This module generates the location of pieces in the image and the coordinate of them in the chess board. The location of pieces in the image guides the robot to control its arm in order to move the pieces. The coordinate of pieces in the chess board helps the system to understand the game and decide the next moves
Related source code
Face tracking for smart phones
A real-time face tracking system using the mobile camera is proposed
The system can adapt with abrupt movements of the mobile phone and the tracked face
High accuracy is achieved
4D (3D and temporal) micro and macro facial expression recognition (FER)
The facial expression, as one of the nonverbal communications, is the movements and location of facial muscles under the skin. The analysis and measurement of facial expression are essential in understanding the thinking of other people and help us delivering corresponding responses.
Due to difficulties in 2D FER, the research community starts to pay attention to 3D FER. There has been an emergence of several 3D facial expression datasets and algorithms for reconstructing the 3D face model from 2D images.
This project plan to initiate novel micro and macro expression research by considering both video (2D + time) and 4D data (3D + time), utilizing discriminative features and 3D models of faces, as well as additional temporal data.
Document Image Analysis
Optical music recognition on Smart phones.
An optical music recognition system is developed on Samsung galaxy note 2.
Processed music score images are captured by mobile devices with some degrees of distortion.
The system tries to detect the primitive symbols such as, musical notes, rest, dot, cleft, etc.
The recognized song is played and the music XML is extracted.
Source code for correcting the distortion of music scores.
Trajectory recovery of Korean handwritten characters
Handwriting document recognition is a well-known topic in the field of computer vision and pattern recognition.
In order to improve the recognition accuracy, the online handwriting recognition methods try to incorporate the temporal information of the writing process. The offline handwriting recognition methods, however, lack of this temporal information and depend on only the spatial information of the 2D image.
This problem leads to the attempts to recover the writing order from the 2D image for turning the offline recognition problem into the online one.
Deep learning for document image processing
This project studies deep learning and its application to document images analysis and recognition.
Deep learning is employed to improve the performance of document image binarization, text line detection, word segmentation, separation of touching characters.
Source code for the binarization of degraded document images.
Text line segmentation
Binarization results of the sample document images on H-DIBCO 2016 dataset
Weather forcasting and downscaling
EEG-based motor imagery recognition
As a core modality in non-invasive brain–computer interfaces (BCIs), EEG-based motor imagery recognition is increasingly seeking to exploit cross-modal learning to capture discriminative neural patterns from complex multichannel EEG signals. However, existing methods remain limited in jointly preserving the multi-way structure of EEG signals and bridging the semantic gap in single-trial motor imagery decoding.
A novel multiway-based EEG–text alignment framework for motor imagery decoding that integrates tensor-based EEG representation learning with CLIP-inspired cross-modal seman tic alignment is proposed.
Source code for the model training and validation.