Method for carrying out 256-dimensional binary quantization on 128-dimensional floating point type feature descriptor of HardNet

Abstract

The invention discloses a method for carrying out 256-dimensional binary quantization on 128-dimensional floating point type feature descriptors of HardNet, and belongs to the field of image processing. For an input image, firstly, an SIFT algorithm is adopted to detect feature points of the image, the feature points are taken as a center, the input image is rotated according to a rotation angle calculated by SIFT, a sub-image is intercepted and sent to a HardNet network, and a 128-dimensional floating point type feature descriptor is output. Then, 128 floating-point numbers of the feature descriptors are sorted according to a sequence from small to large, and the positions and the sizes of trisection points of the 128-dimensional floating-point feature descriptors are calculated; and the128-dimensional floating point type feature descriptor is quantized into a 256-dimensional binary feature descriptor for image matching. Compared with SIFT, the matching success rate can be improved to 100%; compared with a HardNet floating point type descriptor, the matching speed of the binary descriptor can be increased by about three times.

Description

technical field [0001] The invention belongs to the field of image processing, in particular to a method for performing 256-dimensional binary quantization on a 128-dimensional floating-point feature descriptor of HardNet. Background technique [0002] When the drone is in a complex environment and GPS cannot meet the navigation requirements safely and stably, visual navigation can assist the positioning of the drone by matching aerial images and electronic maps. Among them, the matching method based on local feature points is widely used because of its feature invariance and robustness. SIFT (scale-invariant feature transform) is a traditional feature point extraction method. First, by constructing a Gaussian pyramid and detecting extreme points in the scale space, feature points that exist at different scales are found. Then, a 128-dimensional floating-point feature descriptor is generated by calculating the gradient direction histogram of the feature point neighborhood. ...

AI Technical Summary

Problems solved by technology

However, when the UAV has high real-time requirements for visual navigation, the floating-point feature descriptor takes a long time to calculate the Euclidean distance

Therefore, HardNet, which outputs 128-dimensional floating-point feature descriptors, is difficult to meet the real-time requirements of UAV visual navigation.

Images