A Method of Anti-rotation HDO Local Feature Description for Object Robust Recognition

Abstract

The invention discloses an anti-rotation HDO local feature description method for target robust recognition, comprising the following steps: setting an image circular detection area; calculating the approximate RGT gradient of the image; calculating the structure tensor, and calculating the The main direction and consistency; the circular image is divided into several circular fan-shaped intervals as cells, and every four adjacent cells are combined into a Block, and the feature vector of each Block is calculated; the feature vector of each Block is combined to form HDO characterization. The present invention introduces the RGT transformation technology and constructs the structure tensor in the circular neighborhood of the constructed pixels, which can ensure that the main direction and consistency of each pixel remain unchanged when the picture is rotated, and then calculate the The main direction histogram, and then adopt the spatial pooling division method of circular sector partition, so that the obtained local features have good discrimination ability and anti-rotation transformation ability.

Description

technical field [0001] The invention relates to a local feature description method, in particular to an anti-rotation HDO local feature description method for target robust recognition. Background technique [0002] Local feature description is widely used in image object detection and recognition. In the past two decades, researchers have proposed a large number of local feature description operators. In order to deal with changes in scale, illumination, contrast, and rotation in images, feature descriptions are usually required to be invariant to illumination, scale, and rotation. The well-known description operators with illumination, scale and rotation invariance include SIFT and SURF operators, but their construction process is relatively complicated and the amount of calculation is large. The Histogram of Gradient Orientation (HOG) proposed by Dalal in 2005, due to the simple calculation, can better capture the shape characteristics of the image, and obtain good detec...

AI Technical Summary

Problems solved by technology

The well-known description operators with illumination, scale and rotation invariance include SIFT and SURF operators, but their construction process is relatively complicated and the amount of calculation is large

The Histogram of Gradient Orientation (HOG) proposed by Dalal in 2005, due to the simple calculation, can better capture the shape characteristics of the image, and obtain good detection performance in face and pedestrian detection, but HOG is in the case of large noise and chaotic image background. Therefore, in order to overcome these weaknesses of HOG, Wonjun Kim proposed a more robust HDO feature in 2014. The HDO feature can robustly describe the local main direction structure information of the image, and has a strong effect on noise and illumination. At the same time, it focuses on the description of the main direction structure, similar to the human visual system, and can also ensure good discrimination in complex scenes, so it can be used for various target detection problems, but the original HDO features are not good. It has rotation invariance, so it limits its popularization and application to a large extent

Of course, the feature description has rotation invariance, which does not guarantee that the feature must have good robust recognition performance. Therefore, it is necessary to consider making the feature description method able to resist the possible rotation transformation in the image under the premise of ensuring the identification ability.

[0004] In the existing anti-rotation transformation local feature description technology, SIFT and SURF operators estimate the local main gradient direction as the reference direction, and then perform appropriate direction calibration to achieve anti-rotation transformation, but there may be ± 20° error due to gradient direction estimation , so it is easy to lead to misjudgment

In order to obtain the rotation-invariant feature mode of LBP feature, the pattern feature with the minimum value is found through the sorting strategy. This processing strategy can also be borrowed in the gradient-based direction histogram, for example, find the maximum gradient direction, and then use this as a benchmark, press Rearrange the gradient direction histogram clockwise or counterclockwise, but because it is susceptible to noise and other factors, it is sometimes difficult to obtain a stable sorting relationship, which will cause trouble for the final detection and recognition

Images