A scale and affine invariant image corner detection method and system based on self-supervised learning

Abstract

This invention discloses a scale- and affine-invariant image corner detection method and system based on self-supervised learning, belonging to the field of image detection. The method includes the following steps: acquiring the original input image and performing first-order Gaussian directional derivative filtering and second-order Gaussian directional derivative filtering respectively to construct a corner feature candidate map; fusing the corner feature candidate map with the original input image through a gating mechanism, and then extracting features through a rotation-equivariant network to obtain a corner response sub-map and a direction estimation sub-map; performing global average pooling on the corner response sub-map and the direction estimation sub-map in the channel dimension and spatial dimension to obtain corresponding direction-aware weights, and then performing weighted fusion to obtain a rotation-invariant corner response map and a pixel-level rotation-equivariant direction histogram, thereby realizing corner detection.

Images