Selected work

Representative Papers

An Image-Based Path Planning Algorithm Using a UAV Equipped with Stereo Vision

Oct 17, 2022Annual Conference of the IEEE Industrial Electronics Society

Binocular-vision UAVs suffer from insufficient path safety in complex 3D terrains due to the inherent lack of depth information in 2D images. Method: This paper proposes an image-driven path planning method integrating stereo vision with multi-feature detection. A disparity map is constructed to recover terrain depth; safe waypoints are automatically extracted via joint detection of edges, line segments, and corners; and ArUco markers enable pose estimation for start/end-point identification and trajectory generation. Contribution/Results: The method innovatively synergizes depth reconstruction with heterogeneous image features for robust waypoint selection. Evaluated in V-REP simulation and on a physical UAV platform, it significantly improves path safety and terrain adaptability compared to conventional A* and PRM algorithms.

4 citationsRead paper

Aerial Image Stitching Using IMU Data from a UAV

Jul 27, 2023International Conference on Image, Vision and Computing

Existing image stitching methods for UAV aerial photography suffer from feature matching failure and inaccurate homography estimation due to large inter-frame displacements, strong rotations, and significant camera pose variations. To address this, we propose an IMU-assisted robust image stitching framework that jointly leverages inertial measurement unit (IMU) motion data and visual features to estimate the UAV’s full six-degree-of-freedom (6-DoF) motion. This enables precise perspective distortion correction and high-accuracy homography computation, substantially reducing reliance on sparse or unstable feature points. Experimental results demonstrate that, compared to conventional vision-only approaches, our method improves stitching accuracy by 23.6% and reduces ghosting artifacts by 41.2% under challenging flight conditions involving large displacements and strong rotations. The proposed approach significantly enhances robustness and stability, offering a reliable technical pathway for large-area, high-resolution aerial mapping.

2 citationsRead paper

Distance Transform Guided Mixup for Alzheimer's Detection

May 28, 2025

Early diagnosis of Alzheimer’s disease (AD) is hindered by severe class imbalance in medical imaging data, substantial inter-site variability in MRI acquisition protocols, and limited sample diversity—collectively undermining model generalizability. To address these challenges, we propose Distance-Transform-guided Hierarchical Mixup (DT-HMix): a novel data augmentation framework that leverages distance transforms to encode geometric priors of brain anatomy; integrates spatial hierarchical decomposition with cross-sample layer-wise mixing; and generates high-fidelity, anatomically coherent, and diverse synthetic MRI volumes while preserving structural integrity. DT-HMix is architecture-agnostic, compatible with both CNNs and Vision Transformers (ViTs). Evaluated on the multi-center ADNI and AIBL datasets, DT-HMix significantly improves model generalization and robustness across sites and scanners, outperforming standard Mixup and CutMix baselines in classification accuracy. This work establishes a generalizable, anatomy-aware data augmentation paradigm for cross-center AD neuroimaging analysis.

1 citationsRead paper

Vision-Based System Identification of a Quadrotor

Jul 27, 2023International Conference on Image, Vision and Computing

High model uncertainty in quadcopter dynamics—stemming from inaccurate calibration of thrust and drag coefficients—severely limits control performance and autonomous decision-making. Method: This paper proposes a novel onboard-vision-based system identification paradigm. Real-time pose and motion features are captured via visual sensors and integrated into a gray-box modeling framework that explicitly embeds physical constraints to mitigate parameter coupling. This significantly improves the identification accuracy of critical aerodynamic parameters, including thrust and drag coefficients. An LQR controller is subsequently designed based on the identified model and validated in closed-loop experiments. Results: Compared to conventional black-box or purely first-principles models, the proposed approach achieves superior model fidelity and enhanced robustness in control response. It establishes a high-confidence dynamical foundation for fault detection and autonomous flight, enabling more reliable real-time decision-making and control synthesis.

1 citationsRead paper

Multi-Scale Structural Features for Continual, Comprehensible Visual Recognition in a Developmental Learning Framework

Jul 28, 2026

This work addresses the limitations of existing machine learning approaches in continual learning—specifically, their inability to effectively reuse prior knowledge, susceptibility to catastrophic forgetting, and lack of model interpretability—by proposing a gradient-free developmental learning framework. The method incrementally constructs an interpretable network with a human-understandable topological structure, processing samples one at a time without requiring task boundaries or replay of historical data. It leverages multi-scale shape representations that integrate edges, contours, and their spatial relationships. Experimental results demonstrate that the proposed approach significantly outperforms current replay-free methods on class-incremental MNIST, achieving accuracy comparable to or exceeding that of replay- or regularization-based baselines, while entirely eliminating catastrophic forgetting and preserving model interpretability.

0 citationsRead paper

Recent publications

Latest Papers

Multi-Scale Structural Features for Continual, Comprehensible Visual Recognition in a Developmental Learning Framework

Jul 28, 2026

This work addresses the limitations of existing machine learning approaches in continual learning—specifically, their inability to effectively reuse prior knowledge, susceptibility to catastrophic forgetting, and lack of model interpretability—by proposing a gradient-free developmental learning framework. The method incrementally constructs an interpretable network with a human-understandable topological structure, processing samples one at a time without requiring task boundaries or replay of historical data. It leverages multi-scale shape representations that integrate edges, contours, and their spatial relationships. Experimental results demonstrate that the proposed approach significantly outperforms current replay-free methods on class-incremental MNIST, achieving accuracy comparable to or exceeding that of replay- or regularization-based baselines, while entirely eliminating catastrophic forgetting and preserving model interpretability.

0 citationsRead paper

Variance-Preserving Orthogonal Selection (VPOS): Greedy Feature Selection via Orthogonal Deflation in PCA Loading Space

Jul 25, 2026

This work addresses the challenge of preserving data variance while avoiding redundancy in unsupervised feature selection by proposing a greedy selection method grounded in a weighted PCA loading space. The approach incorporates a null-space ablation mechanism that, after each selection step, removes the variance direction of already-selected features, thereby enforcing subsequent selections to cover orthogonal subspaces within the covariance structure. The method establishes a theoretical connection to monotone submodular maximization and introduces a reproducible hyperparameter selection criterion based on sensitivity scanning of reconstruction mean squared error (MSE). Experimental results across eight benchmark datasets demonstrate that the proposed method consistently achieves the lowest reconstruction MSE, offers speedups of 10–140× over graph-based baselines, and reduces MSE by 10%–73% compared to standard PCA without ablation at equivalent dimensions.

0 citationsRead paper

Label-Decoupled Style Augmentation for Domain Generalization in Multi-Label Remote Sensing Scene Classification

Jul 14, 2026

This work addresses the limited generalization of multi-label remote sensing scene classification in cross-domain deployment, where existing global style augmentation methods often induce inter-class interference due to their neglect of label-specific discrepancies. To mitigate this, we propose a label-decoupled style augmentation framework that leverages label-aware spatial attention to extract localized feature statistics, enabling independent coefficient mixing for cross-domain samples sharing the same labels, followed by attention-weighted normalization for feature reconstruction. Our approach is the first to integrate label-specific attention into multi-label remote sensing domain generalization, achieving label-wise decoupled style perturbations that effectively prevent category contamination while introducing negligible parameters and preserving the original inference pipeline. On leave-one-domain-out benchmarks built from UCM, AID, and DFC15, our best variant achieves an average mAP of 71.5%, outperforming empirical risk minimization by 5.0 points, the strongest global baseline by 1.3 points, and yielding up to a 7.7-point gain in the most challenging transfer scenario.

0 citationsRead paper

Social-Annotate: Self-Healing Browser Extension to Annotate and Collect Social Media Data

Jul 01, 2026

This study addresses the limitations of traditional social media data collection methods, which are often cumbersome, context-disconnected, and prone to failure due to frequent changes in page structures—thereby compromising ecological validity. To overcome these challenges, this work proposes a self-healing browser extension that enables non-technical researchers to conduct code-free annotation and intervention experiments directly within native platform interfaces. The system innovatively integrates a large language model–driven self-healing mechanism capable of automatically detecting and repairing selectors to support robust capture of dynamic content. It also features a no-code form designer and a cross-platform injection framework. Validated across twelve major social media platforms, the approach significantly reduces data collection and maintenance costs while empowering researchers to carry out high-quality studies efficiently.

0 citationsRead paper

Hate Speech Detection in Turkish and Arabic Languages: A Comprehensive Study

Jun 30, 2026

This study addresses the surge of online hate speech in Turkish and Arabic, particularly concerning sensitive topics such as refugees, religion, and ethnicity. To tackle this challenge, the authors construct the first fine-grained multilingual hate speech dataset encompassing five major themes in Turkish and refugee-related discourse in Arabic. They further propose a unified BERT-based multi-task learning framework that simultaneously performs hate category classification, intensity regression, target identification, and key span localization. The model achieves state-of-the-art performance across multiple subtasks, significantly enhancing both detection accuracy and semantic understanding of hate content in these low-resource languages.

0 citationsRead paper