Noise-Suppression FPN (NS-FPN)

• NS-FPN is an architectural variant of FPN that suppresses high-frequency and spatially distributed noise via frequency-domain purification.
• It employs Low-Frequency Guided Feature Purification and Spiral-Aware Feature Sampling to enhance multi-scale feature fusion for robust infrared detection and segmentation.
• Experimental results demonstrate significant reductions in false alarms and improved IoU, detection probability, and mAP with only marginal increases in parameters and FLOPs.

A Noise-Suppression Feature Pyramid Network (NS-FPN) is an architectural variant of the standard Feature Pyramid Network (FPN) framework, explicitly engineered to suppress both high-frequency and spatially distributed noise during multi-scale feature fusion. NS-FPN has been introduced to enhance the robustness and discriminativeness of convolutional neural network-based detectors and segmenters, particularly in the context of infrared small target detection and segmentation (IRSTDS) amid substantial background clutter and sensor noise. Unlike previous FPN variants that primarily emphasize feature representation enhancement, NS-FPN prioritizes noise removal via frequency-domain purification and structured, content-adaptive multi-scale fusion, yielding significant reductions in false alarms and improved detection accuracy (Yuan et al., 9 Aug 2025).

1. Architectural Overview

NS-FPN retains the canonical top-down pathway and lateral connection paradigm of the standard FPN, but introduces two functionally novel modules: Low-Frequency Guided Feature Purification (LFP) and Spiral-Aware Feature Sampling (SFS). Each 1×1 lateral convolution in the FPN is replaced with an LFP block, while conventional upsampling plus summation operations are replaced by SFS (except at the highest pyramid level, where LFP is used in isolation).

The processing chain is as follows:

• The backbone network (e.g., ResNet) extracts multi-level features ($C_1$ through $C_4$), each of which is channel-reduced via a 1×1 convolution to produce $\{X_1, X_2, X_3, X_4\}$.
• Each $X_i$ undergoes LFP, outputting $X_i'$.
• The top-down pathway fuses pyramid levels by recursively applying SFS: $Y_4=X_4'$, $Y_3=\mathrm{SFS}(X_3', Y_4)$, $Y_2=\mathrm{SFS}(X_2', Y_3)$, $Y_1=\mathrm{SFS}(X_1', Y_2)$.
• The fused pyramidal outputs $\{Y_1, Y_2, Y_3, Y_4\}$ are passed to subsequent detection or mask heads.

Parameter and computational complexities increase marginally compared to standard FPN: total parameters grow from 3.91M to 4.17M, and FLOPs from 6.80G to 7.96G on a 512×512 input (Yuan et al., 9 Aug 2025).

2. Low-Frequency Guided Feature Purification (LFP)

The LFP module is designed to suppress high-frequency noise embedded in feature maps while retaining signal components conducive to robust target detection.

• Frequency Decomposition: Each feature map $C_4$0 is transformed to the frequency domain via Discrete Fourier Transform (DFT), yielding $C_4$1. A learnable low-frequency binary mask $C_4$2 separates low ($C_4$3) and high-frequency ($C_4$4) components.
• Low-Frequency Attention: $C_4$5 is inverse-transformed to the spatial domain, upon which an attention map $C_4$6 is computed using channel-concatenated average and max pooling followed by a two-layer 1×1 convolution and sigmoid activation.
• Guided High-Frequency Suppression: The high-frequency component is modulated elementwise by $C_4$7, followed by a gated Gaussian filter with learnable standard deviation $C_4$8 to further attenuate weak activations:

$C_4$9

• Feature Reconstruction: The purified feature is reconstructed as $\{X_1, X_2, X_3, X_4\}$0; equivalently, DWT/IDWT may be used.

Total additional overhead is ≈0.01M params and ≈0.01G FLOPs per stage. Backpropagation flows through both the low-pass filter and the Gaussian kernel with respect to the primary detection objective (Yuan et al., 9 Aug 2025).

3. Spiral-Aware Feature Sampling (SFS)

The SFS module generalizes top-down pyramid fusion by replacing static upsampling/addition with a learnable, content- and geometry-aware attention mechanism anchored in spiral sampling.

• Sampling Offsets: For each head $\{X_1, X_2, X_3, X_4\}$1 and sampling point $\{X_1, X_2, X_3, X_4\}$2, define spiral offsets in polar coordinates $\{X_1, X_2, X_3, X_4\}$3, map to Cartesian $\{X_1, X_2, X_3, X_4\}$4, and add a global learnable bias $\{X_1, X_2, X_3, X_4\}$5.
• Deformable Feature Extraction: At each spatial location, higher-level features are sampled at spiral-offset neighborhoods via bilinear interpolation.
• Cross-Attention Fusion: Query ($\{X_1, X_2, X_3, X_4\}$6) and key/value ($\{X_1, X_2, X_3, X_4\}$7) inputs undergo layer normalization. Standard multi-head attention is computed:

$\{X_1, X_2, X_3, X_4\}$8

$\{X_1, X_2, X_3, X_4\}$9

with scalar form at each location:

$X_i$0

• Parameter and Compute Costs: SFS introduces ∼0.25M additional parameters, ∼1.15G more FLOPs (Yuan et al., 9 Aug 2025).

4. Integration in IRSTDS Frameworks

NS-FPN is implemented as a modular, plug-in neck compatible with IRSTDS models such as MSHNet and YOLOv8n-p2. For a standard 512×512 input, backbone outputs are processed via LFP and SFS at each scale, yielding pyramidal features input to detection and segmentation heads. The design is readily adaptable to variable backbone architectures and target tasks without modification to core detection or segmentation modules.

5. Experimental Validation and Ablation Studies

Comprehensive evaluation on public IRSTDS datasets corroborates the efficacy of NS-FPN in both detection and segmentation:

• On IRSTD-1K, NS-FPN achieves IoU 69.3%, detection probability $X_i$1 95.2%, and false alarm rate $X_i$2 8.6, compared to baseline FPN's 67.0%, 91.2%, and 13.1, respectively. LFP and SFS modules contribute additively and synergistically, as demonstrated by ablation.
• On NUAA-SIRST, NS-FPN achieves IoU 78.8% with $X_i$3 as low as 1.6.
• On detection with YOLOv8n-p2, mAP$X_i$4 is improved from 85.0% to 86.3%.
• Module-wise complexity increments over FPN are marginal (total 4.17M params, 7.96G FLOPs).

These results indicate substantial reductions in false alarms and improvements in both segmentation and detection metrics, with the bulk of gains attributable to LFP's noise purification and SFS's adaptive feature fusion (Yuan et al., 9 Aug 2025).

6. Theoretical Distinction from Related Architectures

While other noise-aware FPN variants—such as the Octave convolution-based semantic attention FPN (OcSaFPN) (Li et al., 2020) and contrastive loss-regularized NT-FPN (Fang et al., 25 Jul 2025)—also incorporate frequency decomposition and/or attention for robust feature fusion, NS-FPN is unique in its direct modularization of frequency-space purification (LFP) and learnable geometric fusion (SFS), and in its explicit ablation of false alarm suppressive effect. Unlike NT-FPN, which leverages contrastive objectives on embedding spaces, NS-FPN performs concrete frequency- and geometry-based modulation with negligible parameter increase and tightly integrated blocks.

A plausible implication is that NS-FPN's emphasis on pre-fusion frequency filtration rather than post-fusion attention or information regularization enables superior noise resilience without substantial model reparametrization or added supervisory loss components.

7. Pseudocode and Reproducibility

The forward computational graph for NS-FPN, as defined in the original work, is summarized as:

$X_i$5

Core module operations (DFT filtering, attention map computation, gated Gaussian filtering, and deformable spiral sampling) are mathematically specified in the primary reference (Yuan et al., 9 Aug 2025). These explicit formulations and the lightweight nature of the modifications facilitate direct reproduction and further experimentation within the context of IRSTDS or any multi-scale feature learning application.