Small target detection-oriented network and difficult sample mining method

Abstract

The invention discloses a small target detection-oriented network and difficult sample mining method. The method comprises the following steps: step 1, extracting trunk features of a current image byusing a trunk network; 2, constructing a neck network, and generating a feature pyramid; 3, constructing a region generation network model, and generating a region of interest; 4, generating trainingdata based on difficult sample mining; 5, cutting the region features by using a region-of-interest alignment module; and step 6, classifying and positioning a target by using a head network. According to the method, the small target detection precision can be remarkably improved, and only little calculation overhead is increased.

Description

technical field [0001] The present invention relates to the problem of small target detection, and more specifically, relates to a network and difficult sample mining method for small target detection. Background technique [0002] Benefiting from excellent flexibility and portability, UAV aerial photography is widely used in agriculture, film and television, surveying and mapping, surveillance, express delivery, outdoor search and rescue and other fields. The automatic processing and intelligent identification of aerial data has become an urgent need of the industry. As one of the key technologies, UAV target detection has become a current research hotspot. [0003] Although deep learning methods have achieved great success in the field of general object detection, their performance in UAV aerial photography scenes is not satisfactory. Different from ordinary images, UAV aerial images face serious small target challenges: more small targets, lower target average resolutio...

AI Technical Summary

Problems solved by technology

Limited by GPU memory, it is difficult for existing deep learning methods to directly process such huge images

Images