Selected Projects

Selected Projects.

A curated portfolio of computer vision research and engineering projects built from peer-reviewed work, lightweight deep learning, and deployment-aware vision systems.

DFF-UNet Architecture
IEEE Trans. Instrumentation & MeasurementVol. 74 · 2025Segmentation
DFF-UNet

DFF-UNet: A Lightweight Deep Feature Fusion U-Net Model for Skin Lesion Segmentation

RFRE encoder with adaptive 3×3 and 1×1 convolutions, PDCP bridge with parallel dilated convolutions at rates (1,3,5), and BSFD decoder using bottleneck blocks and depthwise separable convolutions.

mIoU 0.7926Dice 0.88430.190M params0.109 GFLOPs1.30ms
97.55% fewer parameters and 99.20% fewer GFLOPs than U-Net. Best on ISIC2018. Validated on 4 public datasets.
IEEE Xplore → View Full Details →
LiteFusionNet Architecture
Advanced Engineering InformaticsVol. 65 · 2025Segmentation
LiteFusionNet

Advancing Road Safety: A Lightweight Feature Fusion Model for Robust Road Crack Segmentation

Lightweight encoder with residual blocks, DGA, and DVM modules; EASPP bridge for multi-scale crack feature extraction; Dual-Level Decoder with 3×3 convolutions and channel attention.

Dice 0.7828mIoU 0.64310.493M params0.397 GFLOPs
Outperforms SegNet (29.44M params) with 98.3% fewer parameters. Best Dice on Crack500, DeepCrack, and RCD.
ScienceDirect → View Full Details →
VDXNet Architecture
IEEE Geoscience & Remote Sensing LettersVol. 22 · 2025Detection
VDXNet

VDXNet: A Novel Lightweight Deep Learning Model for Vehicle Detection With Aerial Images

RxDF module with asymmetric depthwise separable convolutions replaces C3K2; LiteFPP with selective pooling at 5×5 and 9×9 replaces SPPF; CRDown replaces Conv at P5.

mAP 96.3% (UCAS-AOD)mAP 98.9% (UAVDT)1.608M params539 FPS
37.72% fewer parameters than YOLO11n while improving mAP by 0.52%. Best across all four aerial datasets.
IEEE Xplore → View Full Details →
LiteFODNet Architecture
Intelligent Data Analysis · SAGE/IOS2025Detection
LiteFODNet

LiteFODNet: A Lightweight Deep Learning Model for Intelligent Detection of Small Objects in Runway Surveillance Data

CMSP with parallel atrous convolutions replaces SPPF; SCR replaces Conv at P5; FFM for channel-wise calibration; SPA with decoupled height/width attention for small object localization.

mAP@50-95 0.6812.515M params1.3ms
+0.89% mAP vs YOLOv8n with −16.4% parameters and −27.8% inference time.
SAGE → View Full Details →
CigDet · SURRONE Architecture
Expert Systems with Applications · ElsevierVol. 267 · 2025Review & Dataset
CigDet · SURRONE

Deep Learning-Based Smoker Classification and Detection: An Overview and Evaluation

Comprehensive survey introducing CigDet — a novel annotated dataset for cigarette localization (open-source, Mendeley Data). Benchmarks YOLOv1 through YOLO11. Proposes the SURRONE drone-based outdoor surveillance framework.

YOLOv9 mAP 83.50%YOLO11 mAP 81.50%11 YOLO variants557 annotated images
First open-source annotated dataset for cigarette localization. First benchmark of all YOLO variants (v1–v11).
ScienceDirect → View Full Details →
Discuss Computer Vision Work