Feature Fusion Method for Low-Illumination Images

Feature Fusion Method for Low-Illumination Images

Aiming at the problem of inaccurate feature extraction of low illumination images, a method is proposed that fuses Scale Invariant Feature Transform into SuperPoint. Firstly, the low illumination image is light-enhanced. Secondly, SuperPoint and SIFT features are fused at feature map level, changing the deep neural network weight by labeling the prob output of the network with the SIFT of input image as the maximum of the current prob at pixel-level. Finally, the loss function is constructed based on homography transformation, its principle between image pairs is used to realize the constraint on network parameters. The training and evaluation are conducted on ExDark dataset, tests and comparisons are conducted on multiple indicators of SOTA on HPatches common dataset. The experimental results show that our method improves the precision and recall than SuperPoint, and performs well in multiple evaluation indicators.