Reconhecimento e análise de rachaduras a partir de imagens para monitoramento em regiões com atividade sísmica frequente

• Ano de defesa: 2019
• Autor(a) principal: Lima, João Miguel Correia
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alagoas; Brasil; Programa de Pós-Graduação em Informática; UFAL
• Palavras-chave em Português: Visão por computador; Aprendizagem de máquina; Aprendizagem profunda; Sistemas de reconhecimento de padrões; Computer Vision; Machine Learning; Deep Learning; Pattern recognition systems; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://www.repositorio.ufal.br/handle/riufal/7280

In riskareas,especiallythosecausedbyfrequentseismicmovements,itisnecessarytomonitor eventssuchaschangesinthesizeofcracksinbuildingsandotherurbanstructures.This process, especiallyinregionswithoutadequateinfrastructure,iscarriedoutthroughempirical observationsmadebytheoccupantsoftheseareasthemselves.Thisactivity,byitsverynature, results ininaccurateandoutdateddatacollection,giventhetimelapsebetweencollectionand measurement ofchanges,whichshouldbedonebyaspecialist.Inthispaper,wepropose algorithms basedoncomputervision,toautomatethisprocessbyemployingDeepLearning techniques. Thus,itispossibleforthecomputertodetectcracksandproperlymeasurecracksin images senttoresponsibleagenciesbyresidents,providingagilityindetectingpotentialrisks to people’sphysicalintegrity,whileprovidingexpertswithaccuratedataforactioneffective preventive.Wewilluseasthesituationexperiencedbyresidentsofsomeneighborhoodsinthe city ofMaceió,stateofAlagoas,Brazil,whererulersarecurrentlyemployedtoidentifythe advanceorindentationsofcracksinrealestate,whicharephotographeddailyandtheimages sent tosecurityagencies.Inthepresentworkwepresentaprocessdividedintothreesteps:The first istheidentificationofpointsofinterestintheruler,suchasdigits,usingtheYOLO,adeep neural networkarchitecture,thenwepresentanalgorithmtofilterthecorrectdigitsandfinallyto identify thecrackanditswidthbyapplyingimageprocessingalgorithmandcameracalibration. Thus, aftertheexperiments,wewereabletoachieve75,65%accuracywhenusingthedarktext implementation with31convolutionallayers.Therefore,thiswork,besidesbeinglifesaving,is also alowcosttoolforstructuralinspection. Keywords: