Comparing image patches using convolutional neural networks

Abstract

To compare image patches is a core task in many computer vision areas. A number of hand-crafted features have been used to find the most similar position to a given pattern from a target image. However, these approaches still suffer from many limitations in tough environments. In this paper, we propose a data-driven approach with convolutional neural networks(CNNs) for robust matching. We design new CNN architectures to measure similarity of two images and carry out template matching through the trained network. Consequently, we demonstrate that our template matching method achieves the state-of-the-art performance even in real-world environments. Moreover, we show our study to determine the suitable CNN architecture through network visualization.