UIU-Net: U-Net in U-Net for Infrared Small Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Current infrared small object detection methods struggle with feature loss and limited distinguishability due to deep classification backbones.
• Infrared small objects often exhibit extreme brightness variations, complicating contrast information extraction.
• Existing methods face challenges in effectively learning multi-level and multi-scale features crucial for detecting tiny objects.

Purpose of the Study:

• To propose a novel and effective framework, UIU-Net, for enhanced infrared small object detection.
• To address the limitations of existing methods in handling feature loss and contrast variations in infrared imagery.
• To improve the multi-level and multi-scale representation learning for small objects.

Main Methods:

• Introduced a "U-Net in U-Net" (UIU-Net) framework, embedding a smaller U-Net within a larger backbone.
• Developed a resolution-maintenance deep supervision (RM-DS) module using Residual U-blocks for multi-scale feature learning and global context.
• Integrated an interactive-cross attention (IC-A) module to encode local context between low-level details and high-level semantic features.

Main Results:

• UIU-Net demonstrated superior performance compared to state-of-the-art methods on SIRST and Synthetic infrared datasets.
• The framework effectively enhances both global and local contrast information for improved detection.
• Achieved strong generalization capabilities on video sequence infrared small object datasets like ATR.

Conclusions:

• UIU-Net offers a simple yet effective solution for infrared small object detection, overcoming limitations of current approaches.
• The proposed architecture successfully learns multi-level and multi-scale representations, crucial for tiny object identification.
• UIU-Net shows significant potential for real-world applications requiring robust infrared small object detection.