Efficient-Lightweight YOLO: Improving Small Object Detection in YOLO for Aerial Images

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Aerial image object detection faces challenges with small, dense objects and limited semantic information.
• Existing detectors are often too large for low-end GPUs, hindering practical application.
• There is a need for efficient and accurate object detection models tailored for aerial imagery.

Purpose of the Study:

• To propose an efficient-lightweight You Only Look Once (EL-YOLO) model for aerial image object detection.
• To improve the accuracy of detecting small and densely distributed objects in aerial scenes.
• To develop a model suitable for low-end GPU environments.

Main Methods:

• Designed and scrutinized three model architectures to enhance focus on small objects.
• Introduced efficient spatial pyramid pooling (ESPP) to augment small-object feature representation.
• Developed the alpha-complete intersection over union (α-CIoU) loss function to address sample imbalance.

Main Results:

• EL-YOLO demonstrated strong generalization and robustness for small-object detection in aerial images.
• With model parameters under 10 million and input size 640x640, EL-YOLOv5 achieved APs of 10.8% and 10.7% on DIOR and VisDrone datasets.
• EL-YOLOv5 improved APs by 1.9% and 2.2% compared to YOLOv5 on these datasets.

Conclusions:

• EL-YOLO effectively overcomes limitations of existing detectors for aerial image object detection.
• The proposed model offers a viable solution for accurate and efficient small-object detection on resource-constrained hardware.
• EL-YOLO presents a significant advancement in aerial image analysis, particularly for detecting small objects.