Towards Efficient Detection for Small Objects via Attention-Guided Detection Network and Data Augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Small object detection in Unmanned Aerial Vehicle (UAV) aerial imagery presents significant challenges due to low resolution and dense object distribution.
• Existing object detection methods often struggle with accurately identifying and localizing small objects with limited features.

Purpose of the Study:

• To improve the performance of small object detection in UAV aerial images.
• To address the challenges posed by small and densely packed objects in aerial imagery.

Main Methods:

• Implemented a data augmentation strategy using image division to increase the number of small objects for training.
• Utilized an image pyramid mechanism with up-sampling and fusion of results from multiple detectors to handle dense objects.
• Integrated an attention mechanism and an additional detection head into the YOLOv5 network architecture to enhance focus on small objects.

Main Results:

• The proposed method significantly improved small object detection performance on the Visdrone2019 and DOTA datasets.
• Combined data augmentation and network structure improvements led to enhanced accuracy in detecting small and dense objects.
• The attention mechanism and added detection head effectively improved the network's ability to capture features of small objects.

Conclusions:

• The developed approach effectively enhances small object detection in challenging UAV aerial image scenarios.
• The integration of data augmentation techniques and a refined network architecture offers a robust solution for small object detection.
• This research contributes a valuable method for improving the accuracy and reliability of object detection in aerial surveillance and analysis.