An Anchor Point Generation Method Based on Geometric Attributes

Abstract

The invention discloses a method for generating anchor points based on geometric attributes, which belongs to the field of deep learning. The method steps include: step 1, extracting the coordinates of all marked bounding boxes in the image, and normalizing the width and height of the marked bounding boxes ; Step 2, clustering the aspect ratio of the marked bounding box by Euclidean distance; Step 3, clustering the size of the marked bounding box by SIoU distance; Step 4, taking the size and aspect ratio cluster center as the interval points, anchor points that generate size and aspect ratio vectors. The invention clusters the aspect ratio and size of the marked bounding boxes by normalizing the marked bounding boxes, and the generated anchor points contain the geometric attributes of the target, and can extract region suggestions to improve the accuracy of target detection.

Description

technical field [0001] The invention relates to the field of target detection, in particular to an anchor point generation method based on geometric attributes. Background technique [0002] With the rapid development of deep learning, models based on convolutional neural networks are widely used in the detection of objects in scenes. Generally, target detection methods based on convolutional neural networks can be divided into two-stage detection methods and one-stage detection methods. The two-stage detection method first extracts the region proposal, then outputs the region proposal to the detector and predicts the target position and category in the region, such as Faster RCNN (Ren Shaoqing et al., Faster RCNN:towards real-time object detection with region proposal networks, IEEE Trans on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149); the one-stage detection method does not need to extract region proposals, and directly extracts the candidate region...

AI Technical Summary

Problems solved by technology

Although deep models such as Faster RCNN and YOLO have good applicability to the detection of general targets, they do not consider prior knowledge such as target attributes, making them still have the problem of unsatisfactory detection accuracy in the application of specific scenarios.

Images