RSFNet: A retention-based network with spatial-frequency joint enhancement for infrared small target detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Infrared Small Target Detection (ISTD) is vital for surveillance but challenged by weak signals and lack of texture.
• Current deep learning methods struggle with long-range dependencies and feature oversmoothing in ISTD.
• Existing approaches often fail to achieve high detection accuracy due to these limitations.

Purpose of the Study:

• To develop an advanced deep learning network, RSFNet, for improved Infrared Small Target Detection (ISTD).
• To address the challenges of weak target signals, limited texture, and feature oversmoothing in ISTD.
• To enhance the balance between modeling long-range dependencies and preserving feature details.

Main Methods:

• Proposed RSFNet, a retention-based network integrating a bidirectional 2D decay-retention attention mechanism within the Vision Transformer (ViT) framework.
• Introduced a Spatial-Frequency Joint Enhancement Module (SFE) to integrate spatial and frequency domain features for improved feature transfer.
• SFE facilitates global-local information interaction to enhance salient target feature representation.

Main Results:

• RSFNet demonstrated superior performance compared to state-of-the-art (SOTA) methods on multiple ISTD datasets.
• The proposed network achieved significantly higher detection accuracy in challenging ISTD scenarios.
• RSFNet achieved a notable reduction in training time, by nearly 32%, indicating improved training efficiency.

Conclusions:

• RSFNet effectively suppresses background noise and captures long-range dependencies for robust ISTD.
• The Spatial-Frequency Joint Enhancement Module (SFE) significantly aids in transferring salient target features, improving detection.
• RSFNet offers a promising solution for ISTD, outperforming existing methods in accuracy and efficiency.