A Novel Deep Learning Framework For Image KeyPoint Description

Many major computer vision tasks, such as image registration, data fusion, and change detection, begin with key-point detection and description. Recently, data-driven methods have challenged many classical approaches in many fields of computer vision. To this end, a variety of deep-learning-based techniques that particularly account for keypoint detection and description have been developed. Convolutional layers are an essential component of many deep-learning-based computer vision applications, but they are not totally rotation invariant. This restricts the capability of a CNN to extract suitable descriptors for keypoints. This paper provides a deep learning framework for image keypoint description to overcome this issue. The proposed method enriches the existing extracted descriptors by CNN. In this approach, a pre-trained network is employed for keypoint detection and descriptor extraction. To introduce a richer descriptor, the rotated versions of the input image are applied and a new descriptor is presented. The evaluation results show that the proposed method outperforms the original CNN framework in terms of the number of accurate correspondences, the proportion of correct correspondences, and the matching error.