Illumination-Invariant Feature Detection based on Shallow Hybrid Network Model

Feature detection is the fundamental problem of computer vision, and it is the foundation of vision tasks such as image processing, image analysis, and image understanding. It is widely used in research fields such as object detection and classification, target tracking, and motion estimation. The research and development of feature detectors with illumination robustness have very important theoretical and application value. The traditional feature-based detection method represented by SIFT solves most geometric transformation problems, while the learning-based detection method represented by TILDE had created a precedent for data-driven feature detection, and has partial illumination robustness. In this paper, we will propose an illumination-robust feature point detection method based on convolutional neural network. This method is based on the Key.Net network and has been further enhanced in terms of illumination robustness.