SFGNet: Semantic and Frequency Guided Network for Camouflaged Object Detection

Dezhen Wang, Haixiang Zhao, Xiang Shen, Sheng MiaoarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces SFGNet, a novel network for detecting camouflaged objects by integrating semantic understanding from text prompts with fine-grained frequency domain analysis. By combining these two crucial aspects, SFGNet significantly improves the segmentation of objects that blend seamlessly into their surroundings, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Camouflaged object detection (COD) is a challenging computer vision task that aims to segment objects that are expertly hidden within their environments, mimicking textures, colors, and shapes. While deep learning models have made strides, current approaches often struggle with two key limitations:

  1. Overlooking Semantic Nuances: Existing methods tend to treat all camouflaged objects similarly, failing to leverage the subtle semantic differences between different types of camouflaged targets and their backgrounds. This is particularly problematic when using textual prompts for guidance.
  2. Neglecting Fine-Grained Frequency Features: The intricate details and boundaries of camouflaged objects are often lost when relying solely on spatial features. Fine-grained frequency information, which captures subtle textural patterns, is crucial for distinguishing these objects from their surroundings.

This paper aims to bridge this gap by developing a network that explicitly addresses both semantic understanding and frequency domain analysis for more robust camouflaged object detection.

Methods

SFGNet is built upon a multi-stage architecture that integrates semantic guidance and frequency modeling. Here are the key components:

  • Cross-modal Semantic Guidance (CSG): This module leverages the power of CLIP (Contrastive Language–Image Pre-training) to extract semantic representations from both visual inputs and textual prompts. It uses a Bidirectional Interaction Network (BIN), Bidirectional Cross-Attention (BCA), and Multi-Scale Feature Alignment (MFA) to fuse these multi-modal features, ensuring a rich understanding of the target object and its context.
  • Frequency Modeling: This is a core innovation. The Multi-Band Fourier Module (MBFM) transforms features into the frequency domain and decomposes them into multiple bands (low, mid, and high frequencies). This allows the network to capture distinct structural information at different frequency levels. The Frequency-Spatial Fusion (FSF) module then intelligently combines these frequency-domain features with spatial features using attention mechanisms.
  • Interactive Structure Enhancement Block (ISEB): To further refine the segmentation and ensure structural integrity, the ISEB attends to auxiliary features, helping to preserve clear boundaries and consistent object shapes.
  • Loss Function: A composite loss function is employed, combining Weighted BCE, Weighted IoU, and Cosine Similarity, to jointly optimize for pixel-wise accuracy, regional consistency, and cross-modal alignment.

Datasets

The authors evaluated SFGNet on three publicly available datasets:

  • CAMO: (Camouflaged Object Detection Dataset)
  • COD10K: (Camouflaged Object Detection Dataset 10K)
  • NC4K: (Not explicitly detailed, but implied to be a camouflaged object dataset)

The training set comprises 1,000 images from CAMO and 3,040 from COD10K, totaling 4,040 images. The test set includes 250 from CAMO, 2,026 from COD10K, and 4,121 from NC4K, for a total of 6,397 images.

Key Insights and Ideas

  • Semantic Prompts Matter: The paper highlights that the way you prompt the model significantly impacts performance. A more descriptive prompt like "Camouflaged <class> naturally blending into the surrounding environment" yields better results than simpler prompts, demonstrating the model's ability to interpret nuanced textual guidance.
  • Frequency is Key for Boundaries: By explicitly modeling different frequency bands, SFGNet can better capture the subtle textural variations that define camouflaged objects and their boundaries, which are often missed by purely spatial approaches.
  • Synergy of Semantics and Frequency: The most significant insight is the powerful synergy achieved by combining semantic understanding (from text) with fine-grained frequency analysis. This dual approach allows the network to not only "know" what it's looking for but also "see" the subtle details that reveal its presence.

Key Results

SFGNet demonstrates strong performance across the evaluated datasets:

  • State-of-the-Art Performance: SFGNet significantly outperforms existing state-of-the-art methods on the CAMO and NC4K datasets. It achieves comparable performance with CGCOD on the COD10K dataset.
  • Quantitative Improvements: The paper reports achieving the best S-measure (Sm) and F-measure (F) scores on CAMO and NC4K, and competitive scores on COD10K. For instance, on NC4K, SFGNet achieved an S-measure of 0.912 and an F-measure of 0.884.
  • Ablation Studies: Ablation studies confirm the effectiveness of each proposed module (BIN, MFA, ISEB, MBFM, FSF), showing progressive performance gains as each component is added to the baseline. The MBFM and FSF modules, in particular, contribute significantly to improved accuracy.
  • Qualitative Results: Visual comparisons in Figure 3 show SFGNet producing more accurate and complete segmentations, especially in challenging scenarios with cluttered backgrounds and blurred boundaries.

Limitations and Future Work

  • Limitations: The paper does not explicitly detail limitations. However, as with most deep learning models, performance might be sensitive to the quality and diversity of training data, and generalization to entirely novel camouflage types could be a challenge. The reliance on CLIP for semantic guidance also implies a dependency on the CLIP model's inherent capabilities and biases.
  • Future Work: The authors suggest that the proposed SFGNet framework provides a strong foundation for future research in camouflaged object detection. Further exploration into more sophisticated cross-modal interaction mechanisms and advanced frequency analysis techniques could lead to even greater improvements. The availability of the code suggests potential for community-driven enhancements and applications.

Summary updated

Abstract

Camouflaged object detection (COD) aims to segment objects that blend into their surroundings. However, most existing studies overlook the semantic differences among textual prompts of different targets as well as fine-grained frequency features. In this work, we propose a novel Semantic and Frequency Guided Network (SFGNet), which incorporates semantic prompts and frequency-domain features to capture camouflaged objects and improve boundary perception. We further design Multi-Band Fourier Module(MBFM) to enhance the ability of the network in handling complex backgrounds and blurred boundaries. In addition, we design an Interactive Structure Enhancement Block (ISEB) to ensure structural integrity and boundary details in the predictions. Extensive experiments conducted on three COD benchmark datasets demonstrate that our method significantly outperforms state-of-the-art approaches. The core code of the model is available at the following link: https://github.com/winter794444/SFGNetICASSP2026.
BibTeX
@misc{wang2025sfgnet,
  author = {Dezhen Wang and Haixiang Zhao and Xiang Shen and Sheng Miao},
  title = {{SFGNet: Semantic and Frequency Guided Network for Camouflaged Object Detection}},
  year = {2025},
  eprint = {2509.11539},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.11539},
}
Figure