Content-Based Image Recovery

摘要

We propose an interesting challenge: recovering from aspect-ratio distorted images based on their contents. Given a distorted image, we want to construct a model to predict its original aspect ratio. Since this is a general task, we build a database on top of Pascal VOC datasets. On the base of recent deep convolutional neural networks (CNNs), we present a multi-scale architecture and construct a spatial pooling layer to overcome the problem. By utilizing the multi-level and spatial information, our approach surpasses other methods by a large margin. Towards a better understanding of this task, we also perform detailed studies on experimental results.