YOLOv5s-DSD: An Improved Aerial Image Detection Algorithm Based on YOLOv5s.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Aerial image object detection faces challenges with small targets, dense distributions, and complex backgrounds.
• Existing algorithms struggle to achieve high accuracy in these demanding scenarios.

Purpose of the Study:

• To develop an improved object detection algorithm, YOLOv5s-DSD, specifically for aerial imagery.
• To enhance the detection of small and densely packed objects within complex aerial scenes.

Main Methods:

• Proposed SPDA-C3 structure to minimize information loss and focus on salient features.
• Introduced a novel decoupled head structure (Res-DHead) with an added small object detection head.
• Replaced Non-Maximum Suppression (NMS) with Soft-NMS-CIOU to mitigate issues with dense target distribution.

Main Results:

• YOLOv5s-DSD demonstrated superior performance over state-of-the-art models on the VisDrone2019 dataset.
• Achieved a 17.4% increase in mAP@0.5 and a 16.4% increase in mAP@0.5:0.95 compared to the original YOLOv5s.
• Effectively addressed challenges of small targets, dense distribution, and complex backgrounds.

Conclusions:

• The proposed YOLOv5s-DSD algorithm significantly improves object detection in aerial images.
• The novel architectural components and NMS replacement are effective in enhancing detection accuracy.
• YOLOv5s-DSD represents a substantial advancement for aerial image analysis tasks.