PaperlayerView exact revision on arXiv

Source-linked AI summary

Dynamic High-frequency Convolution for Infrared Small Target Detection

Ruojing Li, Chao Xiao, Qian Yin, Wei An, Nuo Chen, Xinyi Ying, Miao Li, Yingqian Wang

arXiv:2602.02969v2cs.CV

TL;DR

SIRST detection must separate targets from clutter because both occupy overlapping high-frequency regions but have different grayscale variation patterns. The paper introduces DHiF, which generates locally adaptive, zero-centered filter banks and combines them with standard convolution. Across networks and datasets, DHiF improves detection performance, including under salt-and-pepper noise, while remaining a drop-in convolutional replacement; its effectiveness is limited for few-convolution architectures and clutter that closely resembles targets.

  • Problem

    SIRST detection requires distinguishing targets from clutters that share high-frequency components and similar local contrast despite different underlying grayscale patterns.

  • Method

    DHiF dynamically generates local filter banks with zero-centered parameters to model discriminative high-frequency representations and integrates them with standard convolution.

  • Results

    DHiF consistently improves performance across networks and datasets, reducing Fa by 11.33% versus WTConv, increasing Pd by 1.15% versus CDC, and improving IoU by 3.08% versus WTConv under salt-and-pepper noise.

  • Takeaways & Limitations

    DHiF provides a generally adaptable convolutional operator for learning discriminative representations of high-frequency regions in SIRST detection networks.

  • Takeaways & Limitations

    DHiF may be less effective in few-convolution architectures and when local clutter closely resembles the target, where broader context or semantic information is needed.

Abstract

from arXiv · show

Infrared small targets are typically tiny and locally salient, which belong to high-frequency components (HFCs) in images. Single-frame infrared small target (SIRST) detection is challenging, since there are many HFCs along with targets, such as bright corners, broken clouds, and other clutters. Current learning-based methods rely on the powerful capabilities of deep networks, but neglect explicit modeling and discriminative representation learning of various HFCs, which is important to distinguish targets from other HFCs. To address the aforementioned issues, we propose a dynamic high-frequency convolution (DHiF) to translate the discriminative modeling process into the generation of a dynamic local filter bank. Especially, DHiF is sensitive to HFCs, owing to the dynamic parameters of its generated filters being symmetrically adjusted within a zero-centered range according to Fourier transformation properties. Combining with standard convolution operations, DHiF can adaptively and dynamically process different HFC regions and capture their distinctive grayscale variation characteristics for discriminative representation learning. DHiF functions as a drop-in replacement for standard convolution and can be used in arbitrary SIRST detection networks without significant decrease in computational efficiency. To validate the effectiveness of our DHiF, we conducted extensive experiments across different SIRST detection networks on real-scene datasets. Compared to other state-of-the-art convolution operations, DHiF exhibits superior detection performance with promising improvement. Codes are available at https://github.com/TinaLRJ/DHiF.

I. INTRODUCTION

SIRST detection must distinguish tiny targets from clutter because both occupy high-frequency regions but differ in grayscale variation and semantic meaning. The paper proposes DHiF, a dynamically generated high-frequency convolution that integrates into detection networks and improves performance with negligible computational changes.

  • Challenge: SIRST detection is difficult because targets and structural clutters share high-frequency components and locally similar contrast patterns.Targets include localized intensity peaks, while clutters include edges, corners, bright spots, broken clouds, and man-made objects.
  • Limitations of existing methods: Traditional fixed operators rely on handcrafted assumptions and struggle to dynamically distinguish diverse high-frequency components.Examples include local contrast measurement and top-hat transformation.
  • Proposed approach: DHiF enables explicit discriminative modeling of different high-frequency categories through dynamically generated filter banks.Its filters are conditioned on local input features and are designed to learn distinctive grayscale variation characteristics.
  • Integration: DHiF is designed as a drop-in replacement for standard convolution in arbitrary convolutional SIRST detection networks.The operator can be embedded in network encoders and residual structures.
  • Results: Experiments across detection frameworks show improved predictive capability with negligible changes to computational efficiency.The paper reports extensive evaluation using multiple detection frameworks.

A. Importance of HFC Representation Learning

Attribution analysis shows that SIRST networks rely on target neighborhoods and background texture regions, motivating explicit learning of high-frequency representations. DHiF addresses this need by using locally generated filter banks to model region-specific high-frequency information.

  • Importance of HFC representation learning: Attribution analysis identifies target neighborhoods and textured background regions as influential for target predictions.Examples include cloud edges, building contours, and vegetation textures.
  • Importance of HFC representation learning: Existing networks focus on intensity-gradient regions, indicating that high-frequency representation learning is important for SIRST detection.This observation is reported across multiple detection networks.
  • Dynamic modeling: DHiF generates a dynamic local filter bank from local input features so different regions receive specialized high-frequency processing.The method uses machine learning to adapt operators to region-specific high-frequency characteristics.
  • Dynamic modeling: The filtered local features are then processed with convolution to produce the output feature map.The filtering and convolution operations share a sliding window, and the final map contains selectively focused high-frequency representations.

2) Sensitivity to HFCs:

DHiF is made sensitive to high-frequency components by combining differentiation-inspired filtering with zero-centered dynamic kernels. Its locally adaptive filters process regional grayscale variations before standard convolution produces features that selectively emphasize discriminative high-frequency information.

  • Sensitivity to HFCs:: Fourier differentiation motivates incorporating spatial differentiation into filters to enhance high-frequency responses.The differentiation property maps temporal or spatial derivatives to frequency-weighted responses.
  • Sensitivity to HFCs:: Avoiding constant kernels reduces low-frequency responses, so filter parameters should contain positive and negative values with a sum near zero.DHiF restricts generated kernel parameters to an interval centered at zero.
  • Structure of DHiF:: DHiF normalizes each local feature window, projects it into vectors, and maps the vectors through tanh to generate k^2 region-specific filters in [-1, 1].The generated filters encode local grayscale variation characteristics and are dedicated to individual local regions.
  • Structure of DHiF:: The generated filters are applied channel-wise to flattened local features, producing filtered features that record distinctive information from different high-frequency regions.The filtered features are combined with local input features before standard convolutions are applied.
  • Structure of DHiF:: After all local regions are processed, DHiF outputs a feature map that selectively focuses on high-frequency components and preliminarily learns their discriminative representations.DHiF can replace standard convolution and can be embedded into a dynamic high-frequency residual block.

A. Experimental Setup

The experiments evaluate DHiF across real-scene datasets, detection networks, convolution baselines, and feature visualizations. Results show improved detection and adaptability across networks and datasets, with limited computational impact.

  • Datasets and evaluation: Experiments use the IRSDT-1k and NUAA-SIRST real-scene datasets and evaluate IoU, nIoU, Pd, Fa, FPS, and parameter count.The study compares DHiF with CDC, WTConv, PConv, and standard convolutions across CNN-based and CNN-Transformer SIRST networks.
  • Qualitative analysis: Fig. 3 visualizes target, building-corner, and bright-clutter regions, feature maps from standard convolution and DHiF, and local filter banks for three HFCs.The figure compares feature processing in third-level DNANet variants without and with DHiF.
  • Overall results: DHiF-enhanced networks show obvious quantitative and qualitative detection improvements on both datasets.The reported improvements include more accurate segmentations and better target-background distinction.
  • Comparison with convolution baselines: DHiF consistently improves performance across different networks and datasets, whereas CDC, WTConv, and PConv degrade performance in certain networks.On NUAA-SIRST, the other state-of-the-art convolutions cause severe degradation in most networks, while DHiF remains robust.
  • Efficiency: DHiF improves predictive capability with negligible changes to computational efficiency.The experiments assess both detection performance and computational efficiency across multiple detection frameworks.

C. Ablation Study

The ablation study examines where DHiF should replace standard convolutions and uses feature visualizations to explain its discriminative behavior. Hidden encoder levels provide the main gains, while input-layer replacement is unsuitable.

  • DHiF placement: Replacing DNANet encoder residual blocks at levels 1–5 tests DHiF placement across the input layer and progressively deeper feature-processing stages.The study compares detection performance and computational efficiency for DHiF-Res blocks at different encoder levels.
  • Hidden-layer replacement: IoU increases by 5.05%, nIoU by 0.19%, and Pd by 0.95%, while Fa decreases by 3.20 × 10^-5 on average when DHiF is used at levels 2–5.These gains occur with only a negligible impact on computational efficiency.
  • Input-layer replacement: Input-layer DHiF provides no gain and reduces Pd by 0.38%.The authors suggest that adaptive local filters at the input layer may disrupt the raw image distribution and interfere with subsequent feature extraction.
  • Feature visualization: Without DHiF, DNANet produces consistent positive responses in target, building-edge, and bright-clutter regions, causing false alarms.The visualization study replaces residual blocks in DNANet’s second and third levels with DHiF-Res blocks to examine feature differences and generated local filter banks.
  • Mechanism: DHiF combines a region-specific dynamic filter generator with standard convolution to capture distinctive grayscale variations across HFC regions.This design supports discriminative representation learning while serving as a drop-in replacement for standard convolution.

I. RELATED WORK

SIRST research spans traditional, deep-learning, and high-frequency operator approaches, but existing methods motivate explicit modeling of high-frequency information. DHiF extends this direction as a residual-block-compatible operator for discriminative HFC representation learning.

  • Existing approaches: SIRST methods include traditional and deep-learning approaches, while high-frequency operators commonly use gradients or derivatives to extract HFCs.These operators support tasks including edge detection, sharpening, object detection, and segmentation.
  • Existing approaches: Attribution analysis across several detection networks shows that intensity-gradient regions are important for SIRST predictions.The analysis supports learning representations of HFCs in SIRST detection networks.
  • DHiF integration: DHiF can serve as a drop-in replacement for standard convolution and enhances discriminative learning of different HFCs in SIRST networks.This design is presented as compatible with existing network architectures.
  • DHiF integration: DHiF replaces the first standard convolution in a residual block to form a DHiF-Res block.The block combines DHiF with normalization, activation, shortcut, and standard convolution components.

III. ADDITIONAL EXPERIMENTS

The experiments evaluate profile-description and target-discrimination capabilities using overlap, detection, and false-alarm measures. These metrics support visual and quantitative assessment of SIRST methods.

  • Metrics: IoU and nIoU evaluate an algorithm’s profile description ability using interaction and union areas.IoU aggregates areas across test samples, while nIoU averages per-sample overlap ratios.
  • Visual evaluation: Visual comparisons mark targets with red circles and false alarms with yellow circles across different convolution methods.The figures provide qualitative comparisons for multiple scenes.
  • Metrics: Pd and Fa measure an algorithm’s ability to distinguish true targets from false targets.Pd uses true-positive predicted targets relative to total targets, while Fa uses false-positive pixels relative to total pixels.

B. Implementation Details

The study evaluates DHiF on real infrared datasets and examines its integration into DNANet through qualitative comparisons and structural ablations. Results emphasize improved target prediction and false-alarm suppression relative to competing convolutions.

  • Experimental setup: The experiments use IRSDT-1k with 1001 images and NUAA-SIRST with 427 images, using stated training and optimization settings.IRSDT-1k has 800 training and 201 testing images; NUAA-SIRST has 213 training and 214 testing images.
  • Ablation studies: DHiF-Res variants are compared on NUAA-SIRST to study block structure, with additional comparisons covering kernel size and model complexity.The ablations use DNANet as the typical baseline and modify encoder residual blocks.
  • Results: Networks with DHiF suppress false alarms better than networks using CDC, WTConv, or PConv.The comparison covers specially designed convolutions evaluated for SIRST detection.
  • Ablation studies: Placing DHiF before standard convolution yields significantly superior IoU, nIoU, and Pd performance among the tested block structures.The reported interpretation is that standard convolution further processes locally differentiated high-frequency information from DHiF.

2) Kernel size of DHiF:

The additional studies examine kernel size, nonlinear mapping, feature responses, and computational behavior. They support zero-centered parameter mapping and indicate that larger kernels add cost without obvious detection gains.

  • Kernel size of DHiF: DHiF provides substantial performance gains across tested kernel sizes, while larger kernels add computational cost without obvious detection improvement.The paper attributes this result to the difficulty of capturing global information using convolution layers alone.
  • Nonlinear mapping of DHiF: Only DHiF with tanh(·), whose mapping range is strictly zero-centered, delivers an obvious performance gain in the nonlinear-mapping ablation.Other mappings fail to outperform standard convolution.
  • Feature visualization: DHiF feature visualizations compare target, building-corner, and bright-clutter regions with their local outputs and filter banks.The visualizations examine whether local filters respond differently to distinct HFCs.
  • Computational analysis: The complexity comparison measures model runtime and FPS on 512 × 512 images, while the network contains four convolution operations.These measurements assess computational behavior alongside detection performance.
  • Nonlinear mapping of DHiF: Removing or replacing the zero-centered mapping can introduce feature-map interference and degrade network performance.Unrestricted filter parameters may assume extreme values and lose the zero-centered distribution.

4) Model complexity and runtime:

DHiF adds minimal model size while improving robustness and detection performance across convolutional variants. Its main boundary is that local-detail modeling can be insufficient for target-like clutter and may be less effective in architectures with few convolutions.

  • Model complexity and runtime: Replacing four convolutions with DHiF adds only 0.01 MB of parameters while reducing computation efficiency by 6.85%.The authors attribute the performance improvement to modeling local HFCs rather than increasing model parameters.
  • Robustness to salt-and-pepper noise: Under salt-and-pepper noise, DHiF significantly outperforms CDC, WTConv, and PConv.It reduces Fa by 11.33% versus WTConv, increases Pd by 1.15% versus CDC, and improves IoU by 3.08% versus WTConv.
  • Robustness to salt-and-pepper noise: DHiF's noise-robustness results support its ability to model high-frequency components under corrupted inputs.The comparison is explicitly presented as evidence of DHiF's HFC-modeling capability.
  • Limitations: DHiF is best suited to CNN-based and CNN-Transformer architectures that extract features through many convolutional layers.It may be less effective in architectures with few convolutions, such as some Mamba-based methods, when applied at the input layer.
  • Limitations: Because DHiF focuses on local details, it is fundamentally limited when local clutter patterns closely resemble the target.Accurate discrimination in such cases usually requires broader contextual or semantic information from surrounding regions.
Loading 2602.02969v2…