Feature Point Detection and Description Networks Based on Asymmetric Convolution and the Cross-ResolutionImage-Matching Method

Abstract

Image matching can be transformed into the problem of feature point detection and matching of images. The current neural network methods have a weak detection effect on feature points and cannot extract enough sparse and uniform feature points. In order to improve the detection and description ability of feature points, this paper proposes a self-supervised feature point detection and description network based on asymmetric convolution: ACPoint. Specifically, first, feature point pseudolabels are learned from an unlabeled dataset, and pseudolabels are used for supervised learning; then, the learned model is used to update pseudolabels. Through multiple iterations of model training and label updating, high-quality labels and high-accuracy models are obtained adaptively. The asymmetric convolution feature point (ACPoint) network adopts an asymmetric convolution module to simultaneously train three convolution branches to learn more feature information, which uses two one-dimensional convolutions to enhance the backbone of square convolution from both horizontal and vertical directions and improve the representation of local features during inference. Based on the ACPoint network, a cross-resolution image-matching method is proposed. Experiments show that our proposed network model has higher localization accuracy and homography estimation ability on the HPatches dataset.