Feature Point Detection Based on FPN and Attention Mechanism

In order to solve the problem of low repeatability and accuracy in feature point detection in weak texture regions, the SuperPoint model is integrated with feature pyramid FPN, position Attention, and channel Attention in this paper. Utilizing feature pyramid FPN for multi-scale fusion of feature maps to obtain feature maps containing multi-scale information, in order to obtain sufficient feature points and improve the repeatability of feature point detection. By mapping features through positional Attention and channel Attention, we obtain feature dependencies in both spatial and channel dimensions, thereby enhancing feature relationships and improving the accuracy of feature point detection. Finally, the improved SuperPoint feature point detection model was validated on the Hpatches dataset, and the results showed that the improved SuperPoint model improved repeatability by 2.11 %. In addition, when the pixel difference thresholds were 1, 3, and 5, the accuracy increased by 1.49%, 1.68%, and 2.17%.