Object detection method based on improved feature pyramid network based on clustering algorithm

Abstract

The invention discloses a target detection method for improving a characteristic pyramid network based on a clustering algorithm, and the method comprises four key steps: 1, analyzing the geometric characteristics of a detection object by using the clustering algorithm; 2, constructing a basic detection network main body framework; 3, dynamically adjusting the characteristic pyramid network according to the main size of the target; And 4, dynamically setting the size length-width ratio of the detection candidate box according to the appearance characteristics of the target. According to the method, when the detected network model is constructed, the geometric features of the to-be-detected domain objects are fully considered, the network generated through the method improves the training efficiency of the model, shortens the training time, improves the generalization ability of the model, and improves the accuracy when a small target and a multi-scale target are detected.

Description

technical field [0001] The invention relates to the technical field of image target detection, in particular to a target detection method based on a clustering algorithm improved feature pyramid network for target detection of small targets, multi-scale targets, and targets with a large aspect ratio variation range. Background technique [0002] There are various types of information. In particular, information in the form of images is particularly important. Compared with information such as text and audio, images are more intuitive and contain more information. Fully extracting image information is an important direction of future information processing. The traditional way of extracting image information is mainly manual interpretation, which is inefficient. With the development of science and technology, image acquisition capabilities are getting stronger and stronger, and the number of images is increasing exponentially. Artificial methods can no longer adapt to the dev...

AI Technical Summary

Problems solved by technology

[0004] The existing mainstream detection model algorithms have poor detection results when dealing with small targets and multi-scale targets, especially when facing some unconventional size targets

In real life, the aspect ratio of some objects (such as aircraft and vehicles) is mostly between 1:1 and 2:1, and the position of the object can be well located by using a square or fine-tuning the square during detection , so as to successfully detect the target, but for some special targets (such as trains, bridges, runways, etc.), the aspect ratio of the target is often very large, and some targets (such as ships) have a large aspect ratio Inner fluctuations can cover almost all situations between 1:1 and 20:1. For these targets, it is difficult for mainstream detection algorithms to accurately locate the specific position of the target, so the detection effect is not good

Images