Keypoint Detection Based on Deep Learning to Improve Illumination Robustness
Changjiang Jiang, Yuhang Zhang, Qin Wang, Sifan Sun · 2021 China Automation Congress (CAC) · 2021
Towards the problem of extracting suitable feature points in computer vision, this paper presents a learning- based method to detect keypoints with repeatability and reliability when the light intensity changes drastically of weather or climate conditions. In contrast to patch-based neural networks, a fully convolutional neural network structure is designed that can operate on a full-size image after a forward pass to generate detection points and corresponding descriptors. We introduce feature pyramids, FCN network structures, cell boundary drift, and metric learning for boosting the repeatability and light robustness of interest point detection. Our model, when training on the AMOS image dataset, is able to repeatedly detect a richer set of interest points than other conventional corner point detectors and filter out reliable feature points by non-maximum suppression. Through intense experiments and in-depth analysis, the experimental results show that the proposed method greatly improves the light robustness of keypoints detection compared to ORB, FAST, SURF, and SIFT.