FPT Software Image Processing Research

“Research is to see what everybody else has seen, and to think what nobody else has thought” – Albert Szent-Gyorgyi. And now, the Image Processing Team (IPT) is the first team of FPT Software Lab and the “eldest” part of the FPT Software Solution Board. With nearly ten member, the team is looking for new ideas and dealing with many problems that no other division in FPT Software “has thought”.
IPT’s research focuses on the areas of Computer Vision and Machine Learning; and has been involved in different projects for developing advanced algorithms with applications in different domains: embedded, mobile devices, assistive and quality of life, medical diagnostics. The approach is organized within application areas: Biometrics; Human machine interaction; Pattern Recognition and ADAS.

Biometrics discovers new algorithms for face recognition, gender identification. We have proposed a mathematical model that is able to automatically assign an objective gender score to a frontal face with a correlation of up to 0.85 with the human subjective scores.

To better represent face and eyes, our research focuses on finding discriminative features to characterize patterns. Based on the distribution on discriminant features, we proposed to learn probabilistic classifiers to separate eyes and non-eyes.

Multiple classifiers are then combined within AdaBoost (an ensemble learning method) to form a robust and accurate eye blinking and yawning detector. Based on eye blinking frequency and yawning detection we also proposed an efficient drowsiness classifier.

Human machine interaction aims at analyzing human activities, developing novel interactive interface based on hand/fingers gestures recognition, pedestrian detection. Our system includes detecting and recognizing hand gestures via combining shape, local auto-correlation information and multi-class support vector machines (SVM – an accurate classification method). The evaluation shows that the system recognizes one-handed gestures with more than 93% accuracy in real-time. The efficiency of system execution is very satisfactory, and we are encouraged to develop a natural human-machine interaction in the near future.

Pattern recognition is investigating new form of the suppression of Moiré patterns in digital images, optimizing Quantization table in JPEG compression, automatic speech balloon detection for comic e-books and X-ray image understanding.

Advanced Driver Assistance Systems (ADAS). Among the most fascinating capabilities of intelligent beings is the seamless perception and interaction with their environment. Guidance and control of automobiles comprises a comprehensive example for these capabilities. A human driver needs to perceive and understand the automobile’s environment. Based on the understanding of the scene he plans, initiates, supervises, and controls suitable behavior. Driver assistance systems aim to project those capabilities onto artificial systems.

ADAS offer many important research problems to work on, such as development of different types of sensors, processing of sensor information to extract relevant features, analysis and classification of these features to detect and track pedestrians and vehicles, behavior and intent analysis of drivers as well as human factor interfaces. In last three years, our research focuses on questions regarding the relationship between vehicle real-time detection, object tracking, land detection and car-to-car distance estimation.

The number plate recognition (NPR) algorithm has been developed as an important feature for a ADAS system.  By using multiple object detectors as  Haar-like, Local Binary Pattern, Deformable Part Model for both detection and classification, our system can detect cars and motorbikes in heavy traffic conditions with accuracy 95% daytime and 85% for nighttime.

The NPR can handle Vietnamese and Japanese license plates and the system is stable under different camera positions and weather conditions.

Tran Nguyen Ngoc – FPT Software Lab 

Related posts: