LI-YOLO: An Object Detection Algorithm for UAV Aerial Images in Low-Illumination Scenes DOI Creative Commons

Shouyuan Liu,

Hao He, Zhichao Zhang

et al.

Drones, Journal Year: 2024, Volume and Issue: 8(11), P. 653 - 653

Published: Nov. 7, 2024

With the development of unmanned aerial vehicle (UAV) technology, deep learning is becoming more and widely used in object detection UAV images; however, detecting identifying small objects low-illumination scenes still a major challenge. Aiming at problem low brightness, high noise, obscure details images, an algorithm, LI-YOLO (Low-Illumination You Only Look Once), for images proposed. Specifically, feature extraction section, this paper proposes enhancement block (FEB) to realize global receptive field context information through lightweight operations embeds it into C2f module end backbone network alleviate problems noise detail blur caused by illumination with very few parameter costs. In fusion part, aiming improve performance shallow head are added. addition, adaptive spatial structure (ASFF) also introduced, which adaptively fuses from different levels maps optimizing strategy so that can accurately identify locate various scales. The experimental results show mAP50 reaches 76.6% on DroneVehicle dataset 90.8% LLVIP dataset. Compared other current algorithms, improves mAP 50 3.1% 6.9% Experimental proposed algorithm effectively scenes.

Language: Английский

Enhancing infrared and visible image fusion through multiscale Gaussian total variation and adaptive local entropy DOI
Hao Li,

Shengkun Wu,

Lei Deng

et al.

The Visual Computer, Journal Year: 2025, Volume and Issue: unknown

Published: March 11, 2025

Language: Английский

Citations

1

Foggyfuse: Infrared and Visible Image Fusion Method Based on Saturation Line Prior in Foggy Conditions DOI

Shengkun Wu,

Hao Li, Lei Deng

et al.

Published: Jan. 1, 2025

Infrared and visible image fusion is currently a common method to enhance details information. However, under interference of foggy weather or military smoke bombs, the quality both images will be affected, resulting in greatly reduced effect, which turn affects downstream tasks. In view effect fog on image, we innovatively propose architecture based prior saturation line (SLP). This mainly includes three modules: Dehazing Module (DM), Auxiliary Enhancement (AEM), Edge (EEM). The DM optimizes SLP with weighted guided filtering obtain detailed transmission maps images. map obtained used further infrared image. AEM EEM are combined non-subsampled shearlet transform process enhanced capable restoring intricate achieving natural colors hazy environments, enhancing visual Since there few studies this area no relevant datasets, developed an pair dataset (Foggy) conditions for experiments. qualitative quantitative evaluation results demonstrate that proposed outperforms state-of-the-art techniques Foggy dataset.

Language: Английский

Citations

0

Near-Infrared Hyperspectral Target Tracking Based on Background Information and Spectral Position Prediction DOI Creative Commons

Li Wu,

Mengyuan Wang,

Weilin Zhong

et al.

Applied Sciences, Journal Year: 2025, Volume and Issue: 15(8), P. 4275 - 4275

Published: April 12, 2025

In order to address the problems of in-plane rotation and fast motion during near-infrared (NIR) video target tracking, this study explores application capsule networks in NIR proposes a network method based on background information spectral position prediction. First, history frame extraction module is proposed. This performs matching images through average curve groundtruth value makes rough distinction between background. On basis, frames stored as pool for subsequent operations. The proposed routing combines traditional algorithm with information. Specifically, similarity feature space calculated, weight allocation mechanism dynamically adjusted. Thus, discriminative ability strengthened. Finally, prediction locates center search region next by fusing features adjacent current frame. effectively reduces computational complexity improves tracking stability. Experimental evaluations demonstrate that novel framework achieves superior performance compared methods, attaining 70.3% success rate 88.4% accuracy data. Meanwhile, visible spectrum (VIS) data analysis, architecture maintains competitive effectiveness 59.6% 78.8% precision.

Language: Английский

Citations

0

Research and application of smart insole assisted gait recognition technology DOI

Yan Yuan,

Jing Xu,

Foo Say Wei

et al.

The Journal of Supercomputing, Journal Year: 2025, Volume and Issue: 81(6)

Published: April 29, 2025

Language: Английский

Citations

0

FoggyFuse: Infrared and visible image fusion method based on saturation line prior in foggy conditions DOI

Shengkun Wu,

Hao Li, Lei Deng

et al.

Optics & Laser Technology, Journal Year: 2025, Volume and Issue: 190, P. 113075 - 113075

Published: May 15, 2025

Language: Английский

Citations

0

LI-YOLO: An Object Detection Algorithm for UAV Aerial Images in Low-Illumination Scenes DOI Creative Commons

Shouyuan Liu,

Hao He, Zhichao Zhang

et al.

Drones, Journal Year: 2024, Volume and Issue: 8(11), P. 653 - 653

Published: Nov. 7, 2024

With the development of unmanned aerial vehicle (UAV) technology, deep learning is becoming more and widely used in object detection UAV images; however, detecting identifying small objects low-illumination scenes still a major challenge. Aiming at problem low brightness, high noise, obscure details images, an algorithm, LI-YOLO (Low-Illumination You Only Look Once), for images proposed. Specifically, feature extraction section, this paper proposes enhancement block (FEB) to realize global receptive field context information through lightweight operations embeds it into C2f module end backbone network alleviate problems noise detail blur caused by illumination with very few parameter costs. In fusion part, aiming improve performance shallow head are added. addition, adaptive spatial structure (ASFF) also introduced, which adaptively fuses from different levels maps optimizing strategy so that can accurately identify locate various scales. The experimental results show mAP50 reaches 76.6% on DroneVehicle dataset 90.8% LLVIP dataset. Compared other current algorithms, improves mAP 50 3.1% 6.9% Experimental proposed algorithm effectively scenes.

Language: Английский

Citations

0