We propose an end-to-end transformer-based estimator that can predict the geometric transformation parameters of an image. Deviating from the existing classification-based formulation, we provided a more generalized method by directly estimating the transformation matrix. We note that the frequency peak position of the inherent resampling artifacts leaves explicit clues for the geometric transformation. 

Role: Main model and comparison paper implementation (CNN-based models and hand-crafted methods), Dataset collection automation, Paper writing

ACM MM 2022, Published [Paper] [Code]

ISSN: 1947-4598

Developed a novel methodology to understand driver habits using multimodal technology that integrates various data obtained from autonomous vehicles. Analyzed the relationship between Advanced Driver Assistance Systems (ADAS) activation and driver identification and profiling. Aimed to understand and analyze driver reactions in various situations.

Role: Driver profiling model structure design, Main architecture implementation contribution, Validation dataset collection and preprocessing

Development of a multimodal learning model for driver and autopilot identification

The global rise in out-of-hospital cardiac arrests underscores the importance of cardiopulmonary resuscitation (CPR) training. To address this, we propose a deep learning solution that transforms smartphone-captured chest compression videos into images for feedback. This model assesses four key CPR quality indicators: compression count, depth, complete release, and hand positioning. 

Role: Video classification comparison paper implementation, Main model implementation contribution, Key image processing technology development (average image, critical image), Paper writing 

ELSEVIER 2024, Published [Paper]

ISSN: 0957-4174

Analyzed vulnerabilities that reduce the accuracy of Optical Character Recognition (OCR) technology to enhance the security of confidential documents and improve the efficiency of cyber threat response through automated classification and recognition technologies. Developed an AI model for recognizing text within images to automate the classification of cyber threat documents. Built a pipeline integrating layout recognition and OCR technologies.

Role: Layout parser development, Text detection, Text recognition, Report writing and review

Development of an end-to-end pipeline for extracting text from images within documents

The purpose of learning a neural network is to facilitate comprehension of the network structure and enhance its generalization performance. We experimented to explore the relationship between loss landscapes and generalization performance. Utilizing the transformer architecture's success based on the multi-head self-attention (MSA) layer in vision tasks, we aim to establish a flatter loss landscape and observe an enhanced generalization performance.

Role: Main architecture, Sliding Window MSA, SAM optimizer, Loss landscape visualization implementation and application, Paper writing

ACK 2022.11.02. Published

ISSN: 2671-7298