A streamlined approach for intelligent ship object detection using EL-YOLO algorithm

Scientific Reports, Год журнала: 2024, Номер 14(1)

Опубликована: Июль 2, 2024

Abstract Maritime objects frequently exhibit low-quality and insufficient feature information, particularly in complex maritime environments characterized by challenges such as small objects, waves, reflections. This situation poses significant to the development of reliable object detection including strategies loss function understanding capabilities common YOLOv8 (You Only Look Once) detectors. Furthermore, widespread adoption unmanned operation intelligent ships have generated increasing demands on computational efficiency cost hardware, necessitating more lightweight network architectures. study proposes EL-YOLO (Efficient Lightweight You algorithm based YOLOv8, designed specifically for ship detection. incorporates novel features, adequate wise IoU (AWIoU) improved bounding box regression, shortcut multi-fuse neck (SMFN) a comprehensive analysis greedy-driven filter pruning (GDFP) achieve streamlined design. The findings this demonstrate notable advancements both accuracy characteristics across diverse scenarios. exhibits superior performance using RGB cameras, showcasing improvement compared standard models.

Язык: Английский

SOD-YOLOv8—Enhancing YOLOv8 for Small Object Detection in Aerial Imagery and Traffic Scenes

Sensors, Год журнала: 2024, Номер 24(19), С. 6209 - 6209

Опубликована: Сен. 25, 2024

Object detection, as a crucial aspect of computer vision, plays vital role in traffic management, emergency response, autonomous vehicles, and smart cities. Despite the significant advancements object detecting small objects images captured by high-altitude cameras remains challenging, due to factors such size, distance from camera, varied shapes, cluttered backgrounds. To address these challenges, we propose detection YOLOv8 (SOD-YOLOv8), novel model specifically designed for scenarios involving numerous objects. Inspired efficient generalized feature pyramid networks (GFPNs), enhance multi-path fusion within integrate features across different levels, preserving details shallower layers improving accuracy. Additionally, introduce fourth layer effectively utilize high-resolution spatial information. The multi-scale attention module (EMA) C2f-EMA further enhances extraction redistributing weights prioritizing relevant features. We powerful-IoU (PIoU) replacement CIoU, focusing on moderate quality anchor boxes adding penalty based differences between predicted ground truth bounding box corners. This approach simplifies calculations, speeds up convergence, SOD-YOLOv8 significantly improves surpassing widely used models various metrics, without substantially increasing computational cost or latency compared YOLOv8s. Specifically, it increased recall 40.1% 43.9%, precision 51.2% 53.9%, mAP0.5 40.6% 45.1%, mAP0.5:0.95 24% 26.6%. Furthermore, experiments conducted dynamic real-world scenes illustrated SOD-YOLOv8’s enhancements diverse environmental conditions, highlighting its reliability effective capabilities challenging scenarios.

Язык: Английский

MODEL COMPRESSION FOR REAL-TIME OBJECT DETECTION USING RIGOROUS GRADATION PRUNING

iScience, Год журнала: 2024, Номер 28(1), С. 111618 - 111618

Опубликована: Дек. 17, 2024

Язык: Английский