HFGL-Net:Laplacian-Based Explicit High-Frequency Component Guidance for a Multi-Scale Non-local Channel Attention Network in Remote Sensing Object Detection

In this paper, we propose a Laplacian-Based Explicit High-Frequency Component Guidance for a Multi-Scale Non-local Channel Attention Network(HFGL-Net).Specifically, a high-frequency decomposition of the image using the Laplacian pyramid and the introduction of a high-frequency component enhancement model (HCEM) explicitly capture the high-frequency information of the image, providing explicit high-frequency guidance to the network and thus enhancing the network’s detection performance. Furthermore, we propose a Multi-Scale Non-local Channel Attention Mechanism (MS-GLCA), which combines multi-scale and non-local perception mechanisms, adaptively fuses features at different scales, enhances the network’s ability to integrate local features with global dependencies, and more effectively assigns feature weights to the detection network.

Installation

Please refer to install.md for installation and dataset preparation.

Getting Started

This repo is based on yolov5.

Acknowledgements

I have used utility functions from other wonderful open-source projects. Espeicially thank the authors of:

• ultralytics/yolov5.