Uma abordagem de extração de grafite com multiagente e identificação por CNNs

Abstract

Graffitiisanurbaninterventionthatuseswallsandpostsassupportgenerallylinked to a social or political message. These urban representations are often an important social indicator. Mapping and tracking them allows us to understand how these interventions interact with other elements of the urban environment. This work aims to evaluate the use of Convolutional Neural Networks and Multiagent Systems to locate and map graffiti in cities from street level images from Google Street View. The method used was the elaboration of four experiments with the pretrained neural networks and reuse of their classifiers for the new context of identification of Graffiti. We used the fine-tuning technique with images extracted from Flickr and Google Street View. Through the analysis of models, it will be shown that the reutilization of the classifiers is promising, reducing the network training time and getting models with results of 76.9% for the true positive rateon Flickr’s data setand sensitivity 71.43% images in the urban environment. Also, it is worth noting that the multi-agent system can navigate the urban environment of Google Street View and analyze an average of 61 images per minute for each agent.