Bridging Image Enhancement and Object Detection in Underwater Environments: Datasets, Evaluation Frameworks, and Training Strategies

目标检测 计算机科学 人工智能 计算机视觉 假阳性悖论 水下 水准点(测量) 能见度 图像增强 桥接(联网) 模式识别(心理学) 图像质量 对象(语法) 图像(数学) 培训(气象学) 图像处理 注释 异常检测 Viola–Jones对象检测框架 训练集 任务分析 对象类检测 对比度(视觉) 背景减法 视觉对象识别的认知神经科学 对比度增强 特征提取 上下文图像分类 监督学习
作者
Ali Awad
出处
期刊: [Michigan Technological University]
标识
DOI:10.37099/mtu.dc.etdr/2052
摘要

Underwater images suffer from color distortion, low contrast, and reduced visibility due to light absorption and scattering in water. Image enhancement methods aim to correct these degradations, but their effect on downstream tasks such as object detection is poorly understood. This dissertation investigates the relationship between underwater image enhancement and object detection through six interconnected studies presented across the introduction, three core chapters, a consolidated Applications chapter, and conclusions. Two datasets are introduced: CUPDD, a unique underwater plant detection dataset from the Great Lakes region, and RUOD-R, a professional re-annotation of the widely used RUOD benchmark that contains 3.5 times more object instances and over 140 times more small objects than the original. A systematic evaluation of nine enhancement methods shows that quality metrics do not predict detection performance. A unified evaluation framework with per-image analysis reveals that approximately 80% of enhanced images match or outperform originals in detection, while a small fraction of over-enhanced images drives the overall performance drop. The investigation also uncovers that label incompleteness in existing benchmarks causes many valid detections to be counted as false positives, and a flip rate analysis on RUOD finds that 38.1% of detections counted as false positives on the original annotations are valid detections of real objects. Enhancement can also support human labeling and annotation when degraded frames are hard to interpret. Applied work on plant detection, mixed-domain training with enhancement as augmentation, and synthesis of underexposed imagery is summarized in one Applications chapter. The first applied study shows that object detection performance is relatively poor on visually degraded underwater images in CUPDD. In the round goby study, using enhancement as an augmentation during detector training improves detection without changing the detector at inference. The complementary idea of deliberate image degradation, through underexposure synthesis, can help by augmenting training with labeled synthetically underexposed images when real dark scene examples are scarce. A perceptual optimization approach (PDFO) yields synthetic underexposure closer to real low light data than simpler histogram based synthesis.

科研通智能强力驱动
Strongly Powered by AbleSci AI
科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
香蕉觅云的应助被rachel采纳,获得10
刚刚
promise发布了新的文献求助10
刚刚
盛小珊完成签到,获得积分10
1秒前
齐xia完成签到 ,获得积分10
2秒前
William_l_c发布了新的文献求助50
2秒前
ff发布了新的文献求助10
2秒前
Akim的应助被Libra采纳,获得10
2秒前
zhanglx66完成签到 ,获得积分10
3秒前
4秒前
WU哇哇完成签到,获得积分10
4秒前
5秒前
思源的应助被99采纳,获得10
5秒前
rui完成签到 ,获得积分10
7秒前
xyzlancet完成签到,获得积分10
7秒前
直率天德完成签到,获得积分10
7秒前
合适尔风完成签到,获得积分10
8秒前
8秒前
齐xia关注了科研通微信公众号
8秒前
9秒前
10秒前
ww960517完成签到,获得积分10
10秒前
11秒前
12秒前
ning完成签到,获得积分10
13秒前
吃葡萄不吐葡萄皮的应助被YXH采纳,获得30
14秒前
rachel发布了新的文献求助10
14秒前
ssk发布了新的文献求助10
15秒前
16秒前
正直指甲油完成签到,获得积分20
16秒前
往事随风完成签到,获得积分10
16秒前
foxdaopo发布了新的文献求助10
17秒前
William_l_c完成签到,获得积分10
17秒前
万木春完成签到 ,获得积分10
17秒前
MrZ1完成签到,获得积分10
17秒前
只只完成签到,获得积分10
18秒前
11完成签到,获得积分10
19秒前
19秒前
小蘑菇的应助被Mercury采纳,获得10
21秒前
21秒前
李铸润完成签到,获得积分10
21秒前
高分求助中
(应助此贴封号)【重要!!请各用户(尤其是新用户)详细阅读】【科研通的精品贴汇总】 10000
拉伸高分子熔体中缠结行为的模拟和小角中子散射研究 2000
Sex, Aging, and Death in a Medieval Medical Compendium: Trinity College Cambridge MS R.14.52, Its Texts, Language, and Scribe 500
高分子基功能界面水蒸发材料与应用研究 500
外场增强的水汽发电研究 500
基于结构调控的功能化水诱导发电材料与器件研究 500
基于界面调控的高效湿气发电材料与器件研究 500
热门求助领域 (近24小时)
化学 材料科学 医学 生物 纳米技术 计算机科学 工程类 有机化学 化学工程 生物化学 物理 复合材料 内科学 催化作用 光电子学 人工智能 细胞生物学 无机化学 物理化学 数学
热门帖子
关注 科研通微信公众号,转发送积分 8191140
求助须知:如何正确求助?哪些是违规求助? 9578110
关于积分的说明 21856695
捐赠科研通 7790440
什么是DOI,文献DOI怎么找? 3436520
关于科研通互助平台的介绍 2560328
邀请新用户注册赠送积分活动 2483288