Multi-branch detection method for small target detection in remote sensing image

Abstract

The invention provides a multi-branch detection method for small target detection in a remote sensing image. The method comprises the steps that 1, a RepDarkNet backbone network is constructed, feature extraction is conducted on a remote sensing image through the RepDarkNet backbone network, and the RepDarkNet backbone network comprises five feature extraction branches which are connected in series; step 2, constructing a cross-layer fusion network, and performing cross-layer feature fusion on the five feature maps extracted by the RepDarkNet backbone network by using the cross-layer fusion network; and step 3, designing a bounding box regression loss function, and obtaining a small target detection frame based on the bounding box loss function and the fused features. Experiments are carried out on a Dior data set and an NWPU VHR-10 data set, and the small-size target in the remote sensing image can be well detected.

Description

technical field [0001] The invention relates to the technical field of remote sensing target detection, in particular to a multi-branch detection method for small target detection in remote sensing images. Background technique [0002] With the continuous development of advanced aerial technology and remote sensing image sensors, object detection in optical remote sensing images has received more and more research attention. However, due to the inconsistent size of remote sensing objects, most objects are small in size and occupy small pixels, which makes it difficult for existing detection methods to extract features. To detect small objects, the backbone network plays an important role, and a good backbone network can extract more rich feature information. However, in the existing detection methods related to remote sensing, due to the insufficient feature extraction ability of the backbone network, the detection method cannot achieve better performance in remote sensing ...

AI Technical Summary

Problems solved by technology

[0003] Aiming at the problem that the existing detection method has insufficient feature extraction ability for small-sized targets and cannot achieve good results in remote sensing images, the present invention provides a multi-branch detection method for small target detection in remote sensing images

Images