Double-attention remote sensing small target detection method based on FPN and PAN networks

Abstract

The invention relates to a dual-attention remote sensing small target detection method based on FPN and PAN networks, and belongs to the field of computer vision. The method comprises the following steps of: pooling a top-layer feature map in the FPN network to obtain a channel vector, performing matrix operation on the channel vector to obtain a channel attention matrix, normalizing the channel attention matrix to obtain a channel weight matrix, multiplying the weight into the feature map to obtain a feature map with channel weight, and fusing the feature map with low-layer features. In a PAN network, channel compression is firstly carried out on a bottom-layer feature map to obtain a spatial vector, then matrix operation is carried out on the spatial vector to obtain a spatial attention matrix, then the spatial attention matrix is normalized to obtain a spatial weight matrix, the weight is multiplied into the feature map to obtain a feature map with spatial weight, and the feature map with the spatial weight is fused with high-layer features. And finally sending to a detection head to generate a detection result. According to the method, the target detection precision can be enhanced in remote sensing image detection, and interference caused by complex background information in the remote sensing image is overcome.

Description

technical field [0001] The invention relates to a remote sensing small target detection method based on double attention of FPN and PAN network, and belongs to the fields of computer vision and image application. Background technique [0002] Remote sensing small target detection has broad prospects in various fields such as vehicle control and ship scheduling. There are many ways to detect and locate small objects from images captured by satellites or drones. However, the detection performance is not ideal for noisy and low-resolution remote sensing images. Existing detection methods based on deep learning can be roughly divided into two categories: one is a two-stage target detection method represented by Faster R-CNN, which first inputs the feature map output from the backbone into the RPN network (Region Proposal). Network), the input samples are mapped into a probability value and four coordinate values. Accurate candidate regions are obtained through training, and t...

AI Technical Summary

Problems solved by technology

Although these multi-scale detection methods have different degrees of improvement in small target detection performance through different feature fusion methods, they have not been tested in remote sensing images, ignoring the impact of complex background information in remote sensing images. The problem of adverse effects caused by small objects and the problem that the details of small objects in remote sensing images are more likely to be lost

Images