Projects
- github repo links
- reports
- presentations
Biomedical Image Classification based on Meta-Learning
The automated classification and detection of medical images present considerable challenges, primarily stemming from the limited availability of small-scale datasets. State-of-the-art and versatile models typically rely on extensive datasets, which pose substantial demands on both memory and computational resources. In response to this, our study introduces a novel approach to biomedical image classification, leveraging the power of meta-learning. Meta-learning offers a promising avenue to improve classification performance in a cost-effective manner.
Pattern Recognition from Large-Scale Data from Multi-Physics Simulations
In astronomy and astrophysics, computer simulations and big data analysis are vital. While computational efficiency has improved with hardware like GPUs, the surge in data volume from larger, more accurate meshes presents storage and transfer challenges. To address this, real-time data analytics must seamlessly integrate with simulations. The study aims to create a High-Performance Computing (HPC) framework for concurrent pattern recognition and data analysis. This involves identifying relevant patterns on the fly for visualization, storage, and analysis, with a core focus on detecting specific physical attributes in 3D data cubes, like large-scale vortices, potentially linked to phenomena such as Jupiter’s “Great Red Spot.”
Video-based Face Recognition
Video-based face recognition is a complex problem with numerous practical applications like surveillance and access control. It’s more challenging than still image-based recognition due to factors such as larger data volumes, motion blur, low video quality, occlusion, scene changes, and diverse acquisition conditions. This study develops an automatic system that allows users to input the videos taken from surveillance cameras to efficiently detect and recognize human faces that appeared in the scene. The system should also enable the “liveness detection”, i.e., realize if the face is from the real person (not taken from the picture).
Robotized Object Recognition and Pick-and-Place Operations
The study analyzes a combination of automated object detection and grasping algorithms. It aims to solve complications in logistics, manufacturing, and daily life such as management of complex supply chains and warehouses, lack of labor and delivery capacity, detection of product defects, and inconvenience for disabled people. Using Yolov5 for object detection and extracting 3D coordinates from point cloud data, the study integrates Kinova robotic arm grasping via Gazebo simulation with ROS Melodic. The manipulator successfully reached and grasped objects, demonstrating the system’s potential benefits in addressing various real-world issues.
LSTM-based Semantic Analysis and Facial Expression Generation
3D avatars are crucial for conveying information naturally and find applications in customer service, entertainment, and medical HCI. However, animating them can be complex and time-consuming. Extracting emotions from text, especially without facial expressions or voice, is challenging. Additionally, there’s limited research in semantic analysis. This study aims to optimize 3D avatar animation by training an LSTM model with pre-trained GloVe word vectors and animating avatars with emotions using OpenPose for Blender and automated operations in iClone7. It emphasizes the need for accurate emotion detection models rather than focusing solely on 3D animation, given the capabilities of software like iClone7 and Blender with their add-ons and automation for custom models and animations, reducing manual work.
Enhancing and Colorizing Infrared Images in Low Light Conditions
Photography is immensely popular today, from people using smartphones to road tracking algorithms. However, image quality can vary, especially in low-light conditions. While significant features may be detected in low-light images, they often lack clear details and are predominantly dark. To address this, various enhancement techniques to low-light RGB images were applied. Additionally, it was found that Near-infrared (NIR) images can offer superior accuracy compared to RGB. NIR operates outside the human-visible spectrum, providing grayscale images. In this study, the distinction between enhanced RGB images and colorized NIR images in low-light conditions was demonstrated.