Yolov3 anchor frame determination method based on improved k-means clustering

Abstract

The invention provides a Yolov3 anchor frame determination method based on improved k-means clustering. According to the method, a traditional method for directly obtaining the Yolov3 anchor frame through one-time k-means clustering is changed, and two-time clustering is carried out on sample data. In the first clustering, sample data are clustered to three scales with different sizes, and the purpose of scale differentiation is achieved. In the second clustering, on the basis of the three clustering subsets obtained in the previous clustering, clustering is carried out again according to the shape features of the sample labeling box, and finally nine clustering results distributed under three scales are obtained and matched with a Yolov3 anchor box. According to the method, the anchor frame obtained through two times of clustering is more in line with the Yolov3 detection process, the influence of a k-means random initial point on a final result can be reduced, the data size of samples is greatly reduced through two times of clustering, and the algorithm speed is improved.

Description

Technical field [0001] The present invention relates to an anchor frame determination method, and more particularly to an anchor frame based on an improved K-Means clustering YOLOV3 anchor frame, suitable for determining anchor frame of YOLOV3 target detection algorithm. Background technique [0002] In recent years, with the breakthrough development of computer integrity, a variety of goal detection techniques based on deep learning are emerged. As one of the representatives, YOLOV3 is widely used in target tracking, environmental detection, object identification, and more in terms of detection speed and high precision. Starting from YOLOV2, the YOLO series algorithm begins to learn from the Thought of FASTER R-CNN, introducing anchor box to give a rough size of the detection target, avoiding the size and location of blind learning objectives during algorithm training, and improves the model of the model, choose the right The anchor frame can effectively improve the training res...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problem that Yolov3 anchor frames obtained by traditional k-means clustering cannot be reasonably allocated to different detection scales, the present invention provides a method for determining Yolov3 anchor frames based on improved k-means clustering. The present invention proposes two clustering method to get anchor boxes suitable for different detection scales

Images