Optimized small object detection in low resolution infrared images using super resolution and attention based feature...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Infrared (IR) imaging is crucial for object detection in challenging conditions like low light and adverse weather.
• Traditional models struggle with low resolution, noise, and limited details in IR images, hindering performance.
• Small object detection in IR imagery remains a significant challenge due to inherent image quality limitations.

Purpose of the Study:

• To develop an optimized method for small object detection in low-resolution infrared images.
• To improve feature quality and reduce computational complexity for enhanced IR image analysis.
• To address challenges of complex backgrounds, scale variations, and noise in IR object detection.

Main Methods:

• Integration of a lightweight super-resolution network (LightweightSRNet) for enhancing IR image resolution.
• Implementation of a Hybrid Global Multi-Head Attention (HG-MHA) mechanism to focus on targets and reduce noise.
• Development of an improved SC-BiFPN module for effective fusion of multi-level features and a C2f-Ghost-Sobel module for efficient edge extraction.

Main Results:

• Significant improvements in detection performance on the HIT-UAV dataset, with Recall increasing from 70.23% to 80.51%.
• Mean Average Precision (mAP) improved from 77.48% to 83.32%, demonstrating enhanced accuracy.
• The model showed robust performance across multiple datasets, validating its effectiveness for practical IR applications.

Conclusions:

• The proposed approach effectively enhances small object detection in low-resolution infrared images.
• The integration of super-resolution, attention mechanisms, and feature fusion modules leads to superior performance.
• The developed model offers a promising solution for real-world applications requiring accurate IR object detection.