MSRP-TODNet: a multi-scale reinforced region wise analyser for tiny object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Real-time surveillance faces challenges in detecting small, distant objects due to limited pixel data, impacting classifier performance.
• Deep Learning (DL) methods enhance detection via feature maps but often incur high computational costs.

Purpose of the Study:

• To introduce the Multi-Scale Region-wise Pixel Analysis with GAN for Tiny Object Detection (MSRP-TODNet) model.
• To improve the accuracy and efficiency of detecting small objects in real-time surveillance applications.

Main Methods:

• Pre-processing images using Improved Wiener Filter (IWF) and Adjusted Contrast Enhancement Method (ACEM).
• Utilizing Multi-Agent Reinforcement Learning (MARL) for regional pixel analysis and feature map generation.
• Employing an Enhanced Feature Pyramid Network (EFPN) for feature map merging.
• Implementing a Generative Adversarial Network (GAN) for final object detection with bounding boxes.

Main Results:

• MSRP-TODNet achieved a mean Average Precision (mAP) of 84.2% at IoU 0.5 and 54.1% at IoU 0.5:0.95 on the DOTA dataset.
• The model demonstrated superior performance compared to TPH-YOLOv5, YOLOv7-Tiny, and DRDet, with detection performance margins of 1.7%-6.1%.
• Achieved an F1-Score of 84.0%, highlighting its effectiveness in small object detection.

Conclusions:

• MSRP-TODNet offers a robust solution for accurate, real-time small object detection in challenging environments like UAV surveillance.
• The proposed framework effectively addresses the limitations of conventional methods by enhancing feature representation and reducing computational load.
• The model's performance on benchmark datasets validates its potential for practical applications in aerial imagery analysis.