A Texture Reconstructive Downsampling for Multi-Scale Object Detection in UAV Remote-Sensing Images

Area of Science:

• Computer Vision
• Remote Sensing
• Deep Learning

Background:

• Unmanned aerial vehicle (UAV) remote-sensing images pose challenges for object detection due to scale variations and feature loss during downsampling.
• Existing deep networks often lose crucial texture information in multi-scale object detection tasks when employing multi-layer downsampling.

Purpose of the Study:

• To address the degradation in multi-scale object detection performance caused by texture feature loss in UAV remote-sensing images.
• To propose a novel, lightweight Texture Reconstructive Downsampling (TRD) module designed to mitigate feature loss during the downsampling process.

Main Methods:

• Proposed a lightweight Texture Reconstructive Downsampling (TRD) module that models lost texture features as residual information.
• Implemented cascading downsampling and upsampling operators within TRD to provide residual feedback for feature map reconstruction.
• Replaced existing downsampling modules in backbone networks with the TRD module for evaluation.

Main Results:

• The TRD module improved average precision (AP) by 3.1% on the NWPU VHR-10 dataset compared to the baseline.
• On the VisDrone-DET dataset, TRD enhanced AP by 3.2%, with significant improvements in AP for small (APS), medium (APM), and large (APL) objects by 3.1%, 8.8%, and 13.9%, respectively.
• TRD demonstrated enriched feature information post-downsampling with minimal additional computational cost.

Conclusions:

• The proposed TRD module effectively reconstructs lost texture features, enhancing discriminative capabilities for subsequent vision tasks.
• TRD significantly improves the accuracy of multi-scale object detection in challenging UAV remote-sensing image datasets.
• The lightweight design of TRD offers an efficient solution for improving feature extraction in deep networks for remote sensing applications.