人工智能深度学习在考古遗址检测中的应用

版权申诉

141 浏览量更新于2024-06-13 收藏 2.72MB DOCX 举报

本文探讨了利用人工智能深度学习网络在考古遗址检测中的应用潜力。随着科技的进步，考古学正在借助现代技术手段提升研究效率和精确度。标题《利用人工智能深度学习网络进行考古遗址检测》聚焦于将深度学习这一前沿AI技术引入考古工作，以识别、定位和分析古代遗迹。深度学习模型，特别是卷积神经网络（Convolutional Neural Networks, CNN），因其在图像识别和模式分析方面的强大能力，能够处理高维度的数据，如卫星图像、无人机拍摄的遥感资料，以及地面测量数据，如地表地形、地质结构等。在考古领域，表面地形学（Surface Topography）是关键的考察对象，其中的几何计量学和特性分析有助于揭示遗址的潜在位置和特征。例如，论文引用了Alexandra Karamitrou等人2022年的研究，他们通过深度学习网络对地表纹理和形状特征进行分析，以识别考古遗址的细微线索，从而辅助发掘工作。另外，本文还提到了其他技术的应用，如铁和锰氧化物在晚期古玻璃色彩中的作用分析，这依赖于微吸收光谱（Micro-XANES）和微能X射线荧光光谱（Micro-XRF）技术，这些非破坏性测试方法提供了对古代材料化学成分的深入理解。同时，地磁场和地电场探测技术也被用于探测埋藏在Amorium古城（位于土耳其中西部）的遗迹，这是环境地球物理学在考古领域的实际应用。无人机（Unmanned Aerial Vehicle, UAV）数据采集技术则为考古遗址的识别和地图制作提供了高效工具，它能够覆盖大面积区域，获取高分辨率图像，对于快速评估和规划考古项目至关重要。W Handayani等人分享了他们在无人机数据分析方面的研究成果，展示了其在田野考古实践中的价值。值得注意的是，该论文从2022年2月接收，经过修订后于9月23日接受发表，并最终于10月3日发布，显示了科学出版过程的严谨性和及时更新。开放获取政策使得研究人员可以方便地访问和共享这些创新性的研究方法，促进了考古学与信息技术的深度融合，为未来的考古工作开辟了新的可能性。

Surf. Topogr.:

Metrol.

Prop.

(2022) 044001

A Karamitrou et al

imitating human intelligence (e.g., Dey, 2016;

Copeland 2020).

Over the past three decades, applications of

machine learning (ML) methods have seen signiﬁcant

increase in Archaeology. ML algorithms such as sup-

port vector machine (Cortes and Vapnik 1995; Kao

et al 2004) random forests (Ho 1995; Ho 1998),

K-means (Cao et al 2009; Jin and Han, 2011; Qi et al

2017) and other similar approaches have been widely

adopted with considerable success in detecting or clas-

sifying archaeological sites, and artifacts (e.g., Kintigh

and Ammerman 1982; Baxter 2009; Menze and

Ur 2012; Flores et al 2019; Orengo et al 2020). These

methods, often referred to as traditional ML algo-

rithms, require the careful selection of input features

(e.g., various spectral indices in satellite imaging) by

human-experts, that are important for the outcome.

Then through an iterative optimization process by the

input of exemplar data the algorithm is trained

based upon multivariate statistics and progressively

improves its performance. Since it requires the deter-

mination and the prior calculation of a range of possi-

ble statistically signiﬁcant input features, it inevitably

suffers from a level of bias as although the training

procedure can point out which from the features are

statistically insigniﬁcant, it cannot suggest, or extract

features different than the provided ones. Also, the

relatively limited number of the features in most appli-

cations often cannot fully describe the targets at

different situations or environmental conditions.

Therefore, the applicability of these algorithms is often

limited to speciﬁc cases and restricts the identiﬁcation

to features with limited spectral and geometric

variations.

In the early 2000s a new machine learning technol-

ogy emerged known as Deep Learning (DL) based on

Artiﬁcial Neural Networks (ANN), and in the case of

image applications, Convolutional Neural Networks

(CNNs). This new technology was largely based on the

seminal work of Fukushima (1980) as well as Hubel

and Wiesel (1959) that introduced the ‘neocognitron’

(Fukushima 1980; 1983; 2003) and established the use

of convolutional and down-sampling layers. In 1986,

Rina Decher was one of the ﬁrst to use the term ‘deep

learning’ to the machine learning community, in

which ‘deep’ was used to describe the use of multiple

layers in a network. Later, Waibel (1987) proposed the

time delay neural network (TDNN), one of the ﬁrst

convolutional networks followed by LeCun et al

(1989) who applied that in a handwritten character

recognition problem using a 7-level Convolutional

Neural Netowork (CNN), called LeNet-5 (LeCun et al

1998). A signiﬁcant advantage of deep learning meth-

ods is that the feature extraction and selection stage is

performed by the learning algorithm automatically

and not by a person. Yet, this usually requires sig-

niﬁcant amounts of labeled data and considerable

computational resources for the training process. The

utilization of GPUs in the training process was the

turning point for using CNNs in image recognition. In

the 2012 ImageNet competition, the ﬁrst CNN ever

submitted, named AlexNet (Krizhevsky et al 2012),

won the competition. The training of AlexNet used

over one million labeled images about ∼1000 object

categories and took ∼6 days using 2 GPUs (Krizhevsky

et al 2012). Since then, deep neural networks have won

many international pattern recognition competitions

and have attracted broad attention, by outperforming

legacy machine learning methods and handling better

large amounts of data with minimum user interven-

tion (Schmidhuber 2015). As such, they offer con-

siderable potential for archaeology.

Among the common tasks assigned to deep learn-

ing CNN networks are image classiﬁcation, object

detection, and semantic segmentation. Classiﬁcation

is a basic process routinely performed in archaeology

with the objective of classifying groups of images that

share some common features, or objects into one of a

number of predeﬁned classes. For example, AI meth-

ods have been used to analyze use-wear on lithic tools

(e.g., Van den Dries 1998) and to classify and identify

types of pottery (e.g., Hörr et al 2008; Anichini et al,

2021; Pawlowicz and Downum 2021). Caspari and

Crespo (2019), used an object-detection based method

to identify Iron Age burial mounds in aerial imagery.

More recently, Agapiou et al (2021) applied the object

detection method to detect surface ceramics in drone

images. Finally, semantic segmentation algorithms

attempt to analyze images further, by partitioning

them into semantically meaningful parts and after-

wards by classifying each part into one of the ‘X’ pre-

determined classes i.e., interpretable image regions for

instance, archaeological sites, regions of vegetation,

modern structures and others (e.g., Garcia-Garcia et al

2018; Minaee et al 2020). Semantic segmentation

operates at pixel-level in the sense that each pixel of an

image is labeled according to the class it belongs to.

This makes semantic segmentation a much more

complicated and computationally intensive task, yet it

can produce more informative and detailed results

compared to classiﬁcation and object identiﬁcation

(e.g., Kendall et al 2015; Garcia-Garcia et al 2018;

Minaee et al 2020). The value of this approach for

geophysical analysis has been demonstrated in the

work of Küçükdemirci and Sarris’s (2020) using

ground-penetrating radar images.

For all this success, only recently there have been

limited yet increasing work adopting CNN approaches

for the automated detection of archaeological sites

(Trier et al 2018; Caspari and Crespo, 2019; Kazimi

et al 2019; Lambers et al 2019; Rayne et al 2020;

Somrak et al 2020; Soroush et al 2020; Bonhage et al

2021; Verschoof-van der Vaart and Landauer 2021)

from Earth observation (EO) data. In part, this is due

to the need for an abundance of labeled data to enable

the CNN to accurately identify different signatures.

For example, ImageNet, an openly available visual

database designed for use in everyday contemporary

剩余17页未读，继续阅读

百态老人

粉丝: 5119
资源: 2万+

人工智能深度学习在考古遗址检测中的应用

人工智能深度学习.docx

人工智能深度学习

基于深度学习的图像检索.docx

基于深度学习的PCB缺陷检测研究.docx

人工智能论文：基于深度学习的目标检测技术综述.docx

基于深度学习的入侵检测系统综述.docx

PyTorch深度学习实战常用神经网络层.docx

基于生成对抗网络的僵尸网络检测.docx

深度学习word2vec学习笔记.docx

深度学习目标检测方法综述.docx

最新资源