Dual-NMS: A Method for Autonomously Removing False Detection Boxes from Aerial Image Object Detection Results

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing

Background:

• Aerial object detection faces challenges due to top-down perspectives, leading to feature extraction difficulties and numerous false detection boxes.
• Existing post-processing methods primarily address overlapping boxes, struggling to eliminate false positives effectively.

Purpose of the Study:

• To develop a novel post-processing method, dual non-maximum suppression (dual-NMS), to autonomously remove false detection boxes in aerial imagery.
• To propose a new network architecture, the correlation network (CorrNet), to enhance feature extraction capabilities for aerial images.
• To significantly improve the overall performance of object detection in aerial images.

Main Methods:

• Dual non-maximum suppression (dual-NMS) combines detection box density and classification confidence to remove false detections.
• Correlation network (CorrNet) incorporates a correlation calculation layer for feature channel separation and a dilated convolution guidance structure.
• The proposed methods were evaluated on the Vehicle Detection in Aerial Imagery (VEDAI) and Dataset for Object Detection in Aerial Images (DOTA) datasets.

Main Results:

• Dual-NMS autonomously removed over 50% of false detection boxes on VEDAI and DOTA datasets, improving precision while maintaining recall.
• CorrNet demonstrated enhanced feature extraction, leading to a 9.78% increase in mean average precision (mAP) on the DOTA dataset compared to YOLOv3.
• The combined approach of CorrNet and dual-NMS significantly improved object detection performance in aerial images.

Conclusions:

• Dual non-maximum suppression (dual-NMS) is an effective post-processing technique for reducing false positives in aerial object detection.
• The correlation network (CorrNet) architecture enhances feature representation, crucial for accurate detection in challenging aerial perspectives.
• The integration of CorrNet and dual-NMS offers a robust solution for high-performance object detection in aerial imagery.