Natural scene digit classification using convolutional neural networks

Abstract

We used a convolutional neural networks based model to classify scene digits. We proposed the Horizontal and Vertical Feature Block to extract feature from different fields of the input images, which is efficient and has fewer parameters. We introduced a multi-input strategy to add location information to our model, while the traditional methods only use a part of information from the source annotations. More importantly, we released a new dataset for scene digit classification. The new dataset is collected from Baidu street view and mobile photos. The samples in the dataset are from the real world, and they are collected from many kinds of scenes in our daily lives, so that this dataset has huge potential in many applications.