Using convolutional neural nwtwork for signboard detection on street view images

In order to efficiently build and update store information in digital maps, a convolutional neural network (CNN) model called Faster R-CNN proposed in 2015 is used for signboard detection on street view images. Google's Inception-ResNet-v2 model is the feature extractor in our model and a series of fully-connected layers are used for classification and bounding box regression. In the beginning, a portion of street view images is labelled for training model. Then, the bounding boxes and corresponding probabilities of signboard detection results can be obtained by applying our model to the other portion of street view images. In the evaluation, the precision of our method based on CNN is about 94.87%. In additional evaluations, all the precisions are above 93% after respectively adding Gaussian noise, Gaussian blur, horizontal flip, and change of brightness to the testing images, which shows high potential of our model for future applications. For example, the change analysis or character recognition techniques can be applied to street view images acquired by a mobile mapping system for updating store's attribute as well as geographic location automatically.