Infrared ship target detection algorithm PEW_YOLOv8 in complex environments

TL;DR

This paper introduces PEW_YOLOv8, an improved infrared ship detection algorithm that enhances accuracy in complex environments by addressing noise and small target challenges.

Contribution

The novel PEW_YOLOv8 algorithm integrates FFA-Net preprocessing, PGIG-Backbone, EMA-Neck, and WIoU Loss to improve infrared ship detection.

Findings

PEW_YOLOv8 achieves 92.2% detection accuracy on the Raytron Technology infrared ship dataset.

The algorithm increases mAP50 and mAP50:95 by 3.9% and 3.1% compared to YOLOv8.

Abstract

In the infrared ship detection task under complex environments, issues such as high rates of missed and false detections occur due to noise, occlusion, and the indistinct features of small targets. To address these problems, this paper proposes a ship target detection algorithm, PEW_YOLOv8, based on YOLOv8. Firstly, the FFA-Net algorithm is used for image pre-processing to improve the contrast and clarity of the images. Secondly, the PGIG-Backbone network is designed. Through multi-path fusion technology, it ensures that features at different scales can complement each other, enhancing the detail expression ability of small targets. Subsequently, the enhanced multi-scale attention neck network, EMA-Neck, is designed. By means of the attention mechanism, it suppresses background noise, enhances feature information related to the target, and improves the distinguishability between the target and the background. Finally, the WIoU Loss is introduced. Through a more comprehensive method of evaluating bounding boxes, the model can better handle inter-target overlaps, occlusions, and other interferences in complex scenes. Under the same experimental conditions, compared with YOLOv8, the PEW_YOLOv8 algorithm achieves a detection accuracy of 92.2% on the Raytron Technology infrared ship dataset, increasing mAP50 and mAP50:95 by 3.9% and 3.1% respectively.

Figures