AI Meets Infrastructure - The Future of Fast, Accurate Crack Width Measurement

Infra­struc­tu­re inspec­tions today are still slow, expen­si­ve, and most­ly manu­al. Visu­al checks can take hours and often requi­re scaf­fol­ding, night shifts, and spe­cia­li­sed equip­ment. Even then, results are often hete­ro­ge­neous and unre­lia­ble.

As digi­tal tech­no­lo­gies evol­ve, the­re’s no reason their advan­ta­ges should­n’t extend to this field as well. In fact, we’­re alre­a­dy see­ing real-world appli­ca­ti­ons. For exam­p­le, a Japa­ne­se bul­let train ope­ra­tor has adopted an AI-powered sys­tem that replaces the time-con­sum­ing manu­al task of wal­king kilo­me­t­res of track to inspect for loo­se bolts and other issues. This chan­ge has not only enhan­ced accu­ra­cy but also dra­ma­ti­cal­ly impro­ved ope­ra­tio­nal effi­ci­en­cy.

Over the past few months, in the scope of our infra­struc­tu­re pro­jects such as Robo­TUNN, we explo­red this topic in depth, nar­ro­wing our focus to how struc­tu­ral dama­ge, spe­ci­fi­cal­ly crack width, can be mea­su­red auto­ma­ti­cal­ly. Crack width is one of the most important indi­ca­tors of struc­tu­ral degra­da­ti­on. Accu­ra­te­ly deter­mi­ning its seve­ri­ty enables ear­ly inter­ven­ti­on, redu­ces the need for exten­si­ve repairs and mini­mi­zes down­ti­me.

From the wide ran­ge of rese­arch available, we tes­ted three pro­mi­sing methods our­sel­ves. Below, we pre­sent what work­ed well and whe­re chal­lenges remain.

1. Orthogonal Projection (OP) — The Classic Approach

This method forms the back­bone of many image-based infra­struc­tu­re inspec­tion sys­tems. The approach, as used by Wang et al. (2021), com­bi­nes clas­si­cal image pro­ces­sing with deep lear­ning to clas­si­fy crack seve­ri­ty levels based on geo­me­tric width.

To go from geo­me­try to clas­si­fi­ca­ti­on, one then trains a deep lear­ning model to pre­dict seve­ri­ty levels from image input alo­ne, so no manu­al mea­su­re­ments are requi­red after trai­ning.

Strengths:

  • Well-sui­ted for cracks in con­trol­led, struc­tu­red envi­ron­ments like rail tracks.
  • Effi­ci­ent and fast once trai­ned.

Limi­ta­ti­ons:

  • Not ide­al for hea­vi­ly cur­ved or bran­ching cracks as ortho­go­nal direc­tions can beco­me ambi­guous or inac­cu­ra­te.
  • Per­for­mance depends on high-qua­li­ty bina­ry masks; seg­men­ta­ti­on errors redu­ce mea­su­re­ment accu­ra­cy.
  • The ske­le­ton of the crack is extra­c­ted.
  • Irrele­vant bran­ches are pru­n­ed, focu­sing only on the main “trunk” of the crack.
  • Crack inter­sec­tions are remo­ved.
  • Tri­ple-PCA is appli­ed at each ske­le­ton point and aver­a­ged to form a cor­rec­ted crack direc­tion.
  • Final­ly, OP is used to mea­su­re the width bet­ween boun­da­ry points, and the Euclide­an distances are aver­a­ged.
Pictures of orthogonal projection
Figu­re 1 Ortho­go­nal Pro­jec­tion method results

2. OrthoBoundary (OB) Methode - A More Robust Version 

Buil­ding on OP, the OB method, review­ed in Zhe Li et al. (2024), addres­ses some weak­ne­s­ses like noi­sy edges, bran­ching struc­tures, and asym­me­tri­cal cracks. OB inte­gra­tes ske­le­ton-based ana­ly­sis with advan­ced pru­ning, ori­en­ta­ti­on cor­rec­tion, and loca­li­zed PCA to esti­ma­te crack width in com­plex pave­ment images.

  • The ske­le­ton of the crack is extra­c­ted.
  • Irrele­vant bran­ches are pru­n­ed, focu­sing only on the main “trunk” of the crack.
  • Crack inter­sec­tions are remo­ved.
  • Tri­ple-PCA is appli­ed at each ske­le­ton point and aver­a­ged to form a cor­rec­ted crack direc­tion.
  • Final­ly, OP is used to mea­su­re the width bet­ween boun­da­ry points, and the Euclide­an distances are aver­a­ged.

The method out­per­for­med other com­mon­ly used tech­ni­ques, inclu­ding the clas­si­cal OP, in terms of accu­ra­cy, speed, and robust­ness.

Strengths:

  • More robust in noi­sy images.
  • Works well on com­plex crack geo­me­tries, inclu­ding cur­ved and bifur­ca­ted cracks.

Limi­ta­ti­ons:

  • Does not mea­su­re crack width at inter­sec­tions, which are excluded from ana­ly­sis.
  • Still strug­gles with shar­ply cur­ved cracks.
AI Meets Infrastructure - The Future of Fast, Accurate Crack Width Measurement
Figu­re 2 OP and OB Method com­pared.

3. Frequency Domain Analysis — Cracks as Signals

This approach views crack detec­tion through a com­ple­te­ly dif­fe­rent lens. Ins­tead of spa­ti­al­ly mea­su­ring width, it ana­ly­ses the fre­quen­cy cha­rac­te­ristics of cracks using video frames and signal pro­ces­sing tech­ni­ques.

Based on the Cao et al. (2023) paper, this method first appli­es Robust Prin­ci­pal Com­po­nent Ana­ly­sis (RPCA) to decom­po­se video frames into two com­pon­ents: a low-rank matrix repre­sen­ting the back­ground and a spar­se matrix cap­tu­ring crack struc­tures. This sepa­ra­ti­on redu­ces back­ground noi­se and shows fine crack details, wit­hout requi­ring thres­hol­ding or bina­riza­ti­on.

  • A 2D Dis­crete Fou­rier Trans­form (DFT) is appli­ed to the spar­se matrix.
  • The magni­tu­de spec­trum is cent­red, log-trans­for­med, and ana­ly­sed.
  • An equi­va­lent ellip­se is fit­ted to the fre­quen­cy com­pon­ents to extra­ct geo­me­tric fea­tures such as crack width and ori­en­ta­ti­on.
  • A 1D power spec­tral den­si­ty (PSD) cur­ve is also com­pu­ted, track­ing how spec­tral power (i.e., “crack inten­si­ty”) varies with fre­quen­cy over time.

This enables not just width esti­ma­ti­on, but also the crack evo­lu­ti­on under stress. Com­pared to the first two methods which acqui­re direct spa­ti­al mea­su­re­ments of crack geo­me­try, this approach mea­su­res crack cha­rac­te­ristics indi­rect­ly by ana­ly­sing their spec­tral signa­tures over time, making it more sui­ta­ble for dyna­mic moni­to­ring than for pre­cise sta­tic width esti­ma­ti­on.

Strengths:

  • Cap­tures rich crack infor­ma­ti­on bey­ond width, inclu­ding growth trends, ori­en­ta­ti­on, and tem­po­ral evo­lu­ti­on.
  • High­ly resistant to back­ground noi­se due to RPCA prepro­ces­sing.

Limi­ta­ti­ons:

  • Still reli­es on manu­al thres­hol­ding in the fre­quen­cy domain to extra­ct crack fea­tures.
  • Requi­res high-qua­li­ty video input and careful para­me­ter tuning.
AI Meets Infrastructure - The Future of Fast, Accurate Crack Width Measurement
Figu­re 3 Dis­crete Fou­rier trans­form of a crack image:

Major Challenges: Lack of Ground Truth

A signi­fi­cant chall­enge across all methods is the absence of relia­ble ground truth data. Public data­sets are rare­ly anno­ta­ted with actu­al width data. This makes it dif­fi­cult to bench­mark the accu­ra­cy of any method. Fur­ther­mo­re, data hete­ro­gen­ei­ty, such as vary­ing came­ra angles or distances, fur­ther com­pli­ca­tes relia­ble com­pa­ri­son.

Until more com­pre­hen­si­ve data­sets beco­me available, we rely on indi­rect vali­da­ti­on methods and cross-com­pa­ri­sons.

While each method we explo­red has its own strengths and weak­ne­s­ses, they all point in the same direc­tion: crack width mea­su­re­ment can be fas­ter, more relia­ble, and more sca­lable with the help of AI. Manu­al inspec­tion alo­ne is no lon­ger suf­fi­ci­ent, espe­ci­al­ly as infra­struc­tu­re ages and workloads grow. The­se approa­ches show that with the right tools and the right data, we can build smar­ter methods that not only detect pro­blems ear­lier but also help pre­vent them altog­e­ther.

***

Refe­ren­ces

Cao, J., He, H., Zhang, Y., Zhao, W., Yan, Z., & Zhu, H. (2023). Crack detec­tion in ultra­high-per­for­mance con­cre­te using robust prin­ci­pal com­po­nent ana­ly­sis and cha­rac­te­ristic eva­lua­ti­on in the fre­quen­cy domain. Struc­tu­ral Health Moni­to­ring, 23(2), 1013–1024. https://doi.org/10.1177/14759217231178457

Li, Z., Miao, Y., Eskan­da­ri Tor­bag­han, M., Zhang, H., & Zhang, J. (2024). Semi-auto­ma­tic crack width mea­su­re­ment using an Ort­ho­Boun­da­ry algo­rithm. Auto­ma­ti­on in Con­s­truc­tion, 158, 105251. https://doi.org/10.1016/j.autcon.2023.105251

Wang, W., Hu, W., Wang, W., Xu, X., Wang, M., Shi, Y., Qiu, S., & Tut­um­luer, E. (2021). Auto­ma­ted crack seve­ri­ty level detec­tion and clas­si­fi­ca­ti­on for bal­last­less track slab using deep con­vo­lu­tio­nal neu­ral net­work. Auto­ma­ti­on in Con­s­truc­tion, 124, 103484. https://doi.org/10.1016/j.autcon.2020.103484

Categories
Latest News
Your contact

If you have any ques­ti­ons about our pro­ducts or need advice, plea­se do not hesi­ta­te to cont­act us direct­ly.

Tel.: +49 40 22 85 900-0
E-mail: info@statsoft.de

Gui­do Band­holz (Head of Sales)