Infrared ship target detection algorithm PEW_YOLOv8 in complex environments.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Infrared ship detection faces challenges like noise, occlusion, and indistinct small targets, leading to high missed and false detection rates.
• Existing algorithms struggle with complex environments, necessitating improved methods for accurate ship identification.

Purpose of the Study:

• To propose an enhanced ship target detection algorithm, PEW_YOLOv8, based on YOLOv8 for improved performance in complex infrared environments.
• To address limitations in current infrared ship detection methods, particularly concerning small targets and challenging environmental conditions.

Main Methods:

• Image pre-processing using FFA-Net to enhance contrast and clarity.
• A novel PGIG-Backbone network with multi-path fusion for improved small target feature expression.
• An enhanced multi-scale attention neck network (EMA-Neck) to suppress noise and improve target distinguishability.
• Integration of WIoU Loss for better handling of occlusions and overlaps.

Main Results:

• The PEW_YOLOv8 algorithm achieved a detection accuracy of 92.2% on the Raytron Technology infrared ship dataset.
• Demonstrated improvements in mean Average Precision (mAP50) by 3.9% and mAP50:95 by 3.1% compared to standard YOLOv8.
• Successfully enhanced the detail expression ability for small targets and improved distinguishability against background noise.

Conclusions:

• PEW_YOLOv8 offers a significant advancement in infrared ship detection, outperforming standard YOLOv8 in complex scenarios.
• The proposed methods effectively address challenges posed by noise, occlusion, and small targets, leading to more robust and accurate detection.
• This algorithm shows promise for applications requiring reliable ship monitoring in challenging infrared imaging conditions.