Research on low-illumination image feature extraction algorithm based on improved SuperPoint convolutional neural network

SuperPoint algorithm is a deep learning algorithm for feature detection, which has been applied in the field of computer vision to extract interest points in images captured in normal illumination environments. To overcome the difficulties of efficiency and poor performance of traditional SuperPoint algorithm in low illumination situations, an improved strategy is proposed, using the GhostNet v2 model to replace the encoder of the SuperPoint module, and GhostNet v1 to replace the encoder of the MagicPoint module, respectively. A public dataset of MS-COCO images was utilized to perform two MagicPoint and one SuperPoint training process. An experiment conducted on part of the public dataset with low-illumination images shows that the improved SuperPoint network can provide more extracted interest points and achieve higher efficiency compared to traditional SuperPoint models.