Object Recognition in Atmospheric Turbulence Scenes Disen Hu

2025-05-02 0 0 8.27MB 5 页 10玖币

侵权投诉

Object Recognition in Atmospheric Turbulence

Scenes

Disen Hu

Visual Information Laboratory

University of Bristol

Bristol, UK

zr18100@bristol.ac.uk

Nantheera Anantrasirichai

Visual Information Laboratory

University of Bristol

Bristol, UK

n.anantrasirichai@bristol.ac.uk

Abstract—The inﬂuence of atmospheric turbulence on acquired

surveillance imagery poses signiﬁcant challenges in image in-

terpretation and scene analysis. Conventional approaches for

target classiﬁcation and tracking are less effective under such

conditions. While deep-learning-based object detection methods

have shown great success in normal conditions, they cannot be

directly applied to atmospheric turbulence sequences. In this

paper, we propose a novel framework that learns distorted

features to detect and classify object types in turbulent environ-

ments. Speciﬁcally, we utilise deformable convolutions to handle

spatial turbulent displacement. Features are extracted using a

feature pyramid network, and Faster R-CNN is employed as

the object detector. Experimental results on a synthetic VOC

dataset demonstrate that the proposed framework outperforms

the benchmark with a mean Average Precision (mAP) score

exceeding 30%. Additionally, subjective results on real data show

signiﬁcant improvement in performance.

Index Terms—atmospheric turbulence, object detection, deep

learning, deformable convolution, object recognition

I. INTRODUCTION

Atmospheric turbulence consistently degrades visual quality

and has a negative impact on the performance of automated

target recognition and tracking in a scene. These distortions oc-

cur when there is a temperature difference between the ground

and the air, causing rapid upward movement of air layers and

resulting in changes to the interference pattern of light refrac-

tion. This leads to visible ripples and waves in both spatial

and temporal directions in images and videos. Mitigating this

effect is an ill-posed problem with non-stationary distortions

varying across time and space, and a degree of distortions is

unknown. The restoration process is thus complex and time-

consuming. Objects behind the distorting layers are almost

impossible to recognise by machines leading to a failure

of automatic detection and tracking processes. Examples of

applications directly affected with atmospheric turbulence are

video surveillance, security and defence.

Object detection methods on natural and clean images have

been greatly developed achieving high performance in term of

both detection accuracy and computational speed. The state of

the arts are based on deep learning with convolutional neural

networks (CNN) (see recent techniques for object detection

This work was supported by the UKRI MyWorld Strength in Places

Programme (SIPF00006/1).

in [1]). However, the performance of these methods declines

when the features are corrupted by noise or distorted by blur.

This degradation is even more pronounced in the case of at-

mospheric turbulence, where distortions appear randomly and

are spatially and temporally invariant [2]. Up to now, only face

recognition in atmospheric turbulence has been developed [3].

They have also proved that training deblurring and detection

models together gives better results than separating the models.

In this paper, we tackle the multi-class object detection

in atmospheric turbulence scenes without image restoration

process. Our framework is hence fast and straightforward. The

method is developed based on the Faster R-CNN detector [4],

but this should not be limited to. The features are extracted

with a feature pyramid network (FPN) [5], which can deal

with different resolutions, different sizes of the objects, and

different amounts of distortions. To mitigate the effects of at-

mospheric turbulence, we incorporate deformable convolutions

[6], which help reduce the impact of visible ripples along ob-

ject edges caused by atmospheric turbulence. A key contribu-

tion is that, as in the atmospheric turbulent environments, the

objects exhibit visual distortions within small ranges of pixel

displacement appearing randomly at all directions. The use

of deformable convolutions provides ﬂexibility in capturing

the shapes of the objects and assists the FPN in extracting

the appropriate features from the distorted objects. As there

is no ground truth available for this problem, we trained the

model using a synthetic dataset and evaluated its performance

using both synthetic and real datasets. Our code is available

at https://github.com/disen-hu/FasterRcnn FPN DCN.

II. RELATED WORK

Image restoration techniques for atmospheric turbulence

have been extensively studied [7]–[9], with some methods

speciﬁcally addressing moving objects in distorted scenes [2],

[10]–[12]. Deep learning technologies have also gained atten-

tion in atmospheric turbulence mitigation, although they are

still in the early stages of development. Existing architectures

have been employed and retrained using synthetic datasets in

various methods [13]–[15]. Additionally, the use of Complex-

Valued CNN was explored in [16], demonstrating signiﬁcant

improvements over traditional CNN-based approaches.

arXiv:2210.14318v2 [cs.CV] 29 May 2023

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ObjectRecognitioninAtmosphericTurbulenceScenesDisenHuVisualInformationLaboratoryUniversityofBristolBristol,UKzr18100@bristol.ac.ukNantheeraAnantrasirichaiVisualInformationLaboratoryUniversityofBristolBristol,UKn.anantrasirichai@bristol.ac.ukAbstract—Theinfluenceofatmosphericturbulenceonacquiredsurve...

展开>> 收起<<

Object Recognition in Atmospheric Turbulence Scenes Disen Hu.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Object Recognition in Atmospheric Turbulence Scenes Disen Hu

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: