Identificação de danos em veículos sinistrados através de imagens

• Autor(a) principal: José Pedro Lobo Marinho Trocado Moreira
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/107814
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Visual image classification is a research area that involves both computer vision and machinelearning. The task of visually classifying an object consists in assigning an object to a category, orset of categories the object belongs to.Traditionally, visual classification tasks are performed using a two layered system, made upof a first layer featuring an out-of-the-shelf feature extractor and detector, and a second classifierlayer. In most recent years, convolutional neural networks have been shown to outperform suchpreviously used systems.Cars have a paramount role in today's world, and being able to automatically classify damagesin cars is of great interest specially to the car insurance industry. Car insurance companies dealwith car inspections on a daily basis. Such inspections are a manual, lengthy and sometimes faultyprocesses. Processes that bring costs and inconveniences to costumers and insurance companiesalike. Even though the total replacement of such manual inspection processes might still be faraway, developing systems to aid, accelerate or enhance the process might be possible with today'stechnology.There isn't, to my knowledge, much work developed in automatic visual car damage classi-fication, and none of it employs these recent performance improvements in image classificationmade possible through the use of CNNs. This happens in spite of some recent research pointing atthe fact that modern CNN technology does in fact, outperform traditional methods in non damage,car related image classification tasks.I hope to successfully apply state-of-the-art Convolutional Neural Network technology to solvethe problem of automatically identifying, distinguishing and locating damages in car images. Iintend to develop a working prototype of a system that will be able to tell if a given photographexhibits a car with damages or not, and possibly identifying, to some extent, the damaged areaswithin the car. The most promising CNN architectures will be used, taking in account both itsclassification accuracy as well as training and classification times.In order to be able to develop such system, a suitable dataset was gathered. The dataset is veryunbalanced in terms of the represented classes. Such imbalances have important effects that werecorrected with suitable techniques to prevent a significant performance degradation. The datasetis used to both train and measure the performance of the system. Since no car damage datasetsare freely available, the used dataset is composed of images gathered using search engines and carcrash agencies galleries.