A DNN-Based Method for Extracting Promotional Media Elements from Urban Images

Companies and public organizations install billboard advertisements in various parts of urban spaces to promote their products and services. Because these billboards are installed in the real space, it has been pointed out that they cause a landscape pollution. In addition, there are so many billboards in urban spaces that meaningful advertising for consumers are buried. To address these problems, it is conceivable that billboards installed in the real space can be recognized through a camera of a smartphone. For example, automatically erase billboards from the image when we take a picture in an urban, or if advertising is of interest, lead the user to web pages of the product being promoted on billboards. However, to realize them, there are two limitations with the current environment surrounding billboards that must be mitigated. First, there are many billboards, especially in the urban space, and a model has not yet been developed to detect them on a large scale, regardless of their content. Second, an infrastructure to extract the elements on billboards has not been established. In this study, we construct object detection models to solve the first problem. The model targets images and detects them, regardless of their content. For the second problem, we construct extraction models that extract multiple elements from billboards such as 'genre, ' 'advertiser, ' and 'product name.' Finally, we consider characteristics of each model, and present appropriate datasets and methods to construct each model.