TOWARDS TRUSTWORTHY MULTI-LABEL SEWER DEFECT CLASSIFICATION VIA EVIDENTIAL DEEP LEARNING Chenyang Zhao1 Chuanfei Hu1 Hang Shao2 Zhe Wang3 Yongxiong Wang3

2025-05-06 0 0 1.73MB 5 页 10玖币

TOWARDS TRUSTWORTHY MULTI-LABEL SEWER DEFECT CLASSIFICATION VIA

EVIDENTIAL DEEP LEARNING

Chenyang Zhao1, Chuanfei Hu1, Hang Shao2, Zhe Wang3, Yongxiong Wang3

1Key Laboratory of Measurement and Control of CSE Ministry of Education, School of Automation, Southeast University, Nanjing, China

2School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China

3School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, China

{zhaocy, cfhu}@seu.edu.cn, shaohang@njust.edu.cn, 201440049@st.usst.edu.cn, wyxiong@usst.edu.cn

ABSTRACT

An automatic vision-based sewer inspection plays a key

role of sewage system in a modern city. Recent advances fo-

cus on utilizing deep learning model to realize the sewer in-

spection system, beneﬁting from the capability of data-driven

feature representation. However, the inherent uncertainty of

sewer defects is ignored, resulting in the missed detection

of serious unknown sewer defect categories. In this paper,

we propose a trustworthy multi-label sewer defect classiﬁca-

tion (TMSDC) method, which can quantify the uncertainty of

sewer defect prediction via evidential deep learning. Mean-

while, a novel expert base rate assignment (EBRA) is pro-

posed to introduce the expert knowledge for describing reli-

able evidences in practical situations. Experimental results

demonstrate the effectiveness of TMSDC and the superior

capability of uncertainty estimation is achieved on the latest

public benchmark.

Index Terms—Trustworthy visual inspection, evidential

deep Learning, multi-label sewer defect classiﬁcation, sewer

pipelines

1. INTRODUCTION

Underground sewage system is one of the most vital life-

lines in a modern city [1], which can guarantee the commu-

nity health, safety, and manufacture. Vision-based inspection

method is widely applied to maintain the underground sewage

system [2]. The internal situations across the sewer pipes

can be captured via a remote mobile vehicle, while the sewer

inspectors diagnose the defects with a long time of looking

at a screen. Such manual inspection is not only laborsome

and time-consuming, but also may cause ophthalmic diseases

during the high-frequency illumination of the screen. Con-

sequently, how to construct an automatic sewer inspection

method has long been a research topic attracting constant at-

tention in the ﬁeld of sewer inspection [3].

Chenyang Zhao and Chuanfei Hu contributed equally to this work. Cor-

responding author: Chuanfei Hu (cfhu@seu.edu.cn).

Recently, deep learning model has received substantial

interest in industrial applications [4,5]. In the vision-based

sewer inspection community, deep learning also attracts in-

creasing attention from both academia and industry [6,7,8].

Here, we focus on the sewer defect classiﬁcation in the set-

ting of multi-label, in which multiply defect classes in an

image are recognized simultaneously. Although these deep

learning-based methods have achieved acceptable perfor-

mances of sewer defect classiﬁcation, while the inherent un-

certainty of sewer defects might not be considered sufﬁciently

in real-world applications [9]. For instance, some categories

of sewer defects are not appeared from historical data, in

sense that, the trained sewer defect classiﬁcation model has

not seen these unknown defects which is the samples out of

knowledge. Existing deep learning-based methods [6,7,10]

for sewer defect classiﬁcation could not describe the mag-

nitude of epistemic uncertainty across known and unknown

sewer defect categories. The model would be over-conﬁdent

to “trust” the prediction, resulting in the missed detection of

serious unknown sewer defect categories.

In this paper, we propose a trustworthy multi-label sewer

defect classiﬁcation (TMSDC) method for unknown sewer

samples setting. To enable the multi-label sewer defect clas-

siﬁcation model to “know unknown”, we cast the task as an

uncertainty estimation problem via evidential deep learning

(EDL) [11]. EDL describes the uncertainty via a Dirichlet

distribution of class probability, which can be seen as an ev-

idence collection process via a deep neural network. The

collected evidence is leveraged to quantify the uncertainty of

sewer defect prediction, for instance, unknown sewer defect

would present a high uncertainty explicitly. Moreover, we in-

troduce the expert knowledge to model the uncertainty and

propose an expert base rate assignment (EBRA), in which the

realistic base rate can provide reliable diagnosis of sewer de-

fects in practical situations [12]. It is noteworthy that TMSDC

can quantify the uncertainty effectively of model whose capa-

bility of distinguishing the known categories would only be

weakened slightly. The main contributions are summarized

as follows:

arXiv:2210.13782v1 [cs.CV] 25 Oct 2022

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：

TOWARDSTRUSTWORTHYMULTI-LABELSEWERDEFECTCLASSIFICATIONVIAEVIDENTIALDEEPLEARNINGChenyangZhao1,ChuanfeiHu1,HangShao2,ZheWang3,YongxiongWang31KeyLaboratoryofMeasurementandControlofCSEMinistryofEducation,SchoolofAutomation,SoutheastUniversity,Nanjing,China2SchoolofComputerScienceandEngineering,NanjingUn...

展开>> 收起<<

TOWARDS TRUSTWORTHY MULTI-LABEL SEWER DEFECT CLASSIFICATION VIA EVIDENTIAL DEEP LEARNING Chenyang Zhao1 Chuanfei Hu1 Hang Shao2 Zhe Wang3 Yongxiong Wang3.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！

TOWARDS TRUSTWORTHY MULTI-LABEL SEWER DEFECT CLASSIFICATION VIA EVIDENTIAL DEEP LEARNING Chenyang Zhao1 Chuanfei Hu1 Hang Shao2 Zhe Wang3 Yongxiong Wang3

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: