PaperlayerView exact revision on arXiv

Source-linked AI summary

MSRNet: A Multi-Scale Recursive Network for Camouflaged Object Detection

Leena Alghamdi, Muhammad Usman, Hafeez Anwar, Abdul Bais, Saeed Anwar

arXiv:2511.12810v2cs.CVcs.AIeess.IV

TL;DR

Camouflaged object detection is challenging because objects blend into their surroundings, and existing methods still struggle with small and multiple objects in complex scenes. MSRNet combines multi-scale transformer features with recursive-feedback refinement; it reports state-of-the-art results on two benchmarks and second place on two others.

  • Problem

    Camouflaged objects blend with their surroundings, and current methods still struggle in complex scenes, especially with small and multiple objects.

    Source: §1 · pp. 1–2

  • Method

    MSRNet extracts multi-scale features with a Pyramid Vision Transformer, selectively merges them, and recursively refines features using Multi-Granularity Fusion Units and feedback.

    Source: §1 · p. 3

  • Results

    MSRNet achieves state-of-the-art performance on two COD datasets and ranks second on the other two.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The authors conclude that MSRNet addresses COD challenges involving small and multiple camouflaged objects.

    Source: §6 · p. 16

    The authors identify feature-extraction computational efficiency and extending the model to video-based COD as future work.

    Source: §6 · p. 17

Abstract

from arXiv · show

Camouflaged object detection is an emerging and challenging computer vision task that requires identifying and segmenting objects that blend seamlessly into their environments due to high similarity in color, texture, and size. This task is further complicated by low-light conditions, partial occlusion, small object size, intricate background patterns, and multiple objects. While many sophisticated methods have been proposed for this task, current methods still struggle to precisely detect camouflaged objects in complex scenarios, especially with small and multiple objects, indicating room for improvement. We propose a Multi-Scale Recursive Network that extracts multi-scale features using a Pyramid Vision Transformer backbone and combines them with specialized Attention-Based Scale Integration Units, thereby enabling selective feature merging. For more precise object detection, our decoder recursively refines features by incorporating Multi-Granularity Fusion Units. A novel recursive-feedback decoding strategy is developed to enhance the model's understanding of global context, thereby helping it overcome the challenges of this task. By jointly leveraging multi-scale learning and recursive feature optimization, our proposed method achieves performance gains, successfully detecting small and multiple camouflaged objects. Our model achieves state-of-the-art results on two benchmark datasets for camouflaged object detection and ranks second on the remaining two. Our code, model weights, and results are available at https://github.com/linaagh98/MSRNet.

1 Introduction

Camouflaged object detection remains difficult in complex scenes, particularly for small, tiny, and multiple objects. MSRNet combines multi-scale feature extraction with recursive refinement, and reports top-two results across four benchmark datasets.

1 Introduction

  • Figure 1 illustrates three challenging cases: multiple objects, small objects, and tiny objects.These are shown in rows 1, 2, and 3, respectively.Source: Figure 1 · p. 2, §1 · p. 2
  • Existing COD models still struggle in complex scenarios, especially scenes with small or multiple objects, motivating further improvement.Camouflage similarity in color, texture, and size, alongside low light, occlusion, and complex patterns, compounds the challenge.Source: §1 · pp. 1–2
  • MSRNet uses a Pyramid Vision Transformer for multi-scale extraction and Attention-Based Scale Integration Units to selectively merge relevant features.The approach processes images at multiple scales to capture global context and local details.Source: §1 · pp. 2–3
  • Its recursive-feedback decoder carries global context from lower-resolution features to higher-resolution features, while Multi-Granularity Fusion Units refine representations.The paper links this design to detecting multiple objects and small or tiny camouflaged objects.Source: §1 · pp. 2–3
  • MSRNet achieves state-of-the-art performance on two COD datasets and ranks second on the other two, among comparisons with 20 state-of-the-art methods.The introduction identifies the difficult cases as scenes with small and multiple objects.Source: §1 · p. 3

2 Related Works

Related work traces COD from handcrafted features to deep-learning and transformer approaches, with prior methods combining multi-scale aggregation, staged refinement, or joint training. Transformer-based COD methods also differ in how they decode and combine features.

2 Related Works

  • Prior CNN-based COD work is grouped into multi-scale feature aggregation, multi-stage processing, and joint training approaches.Examples include C2FNet and ZoomNet for aggregation, SINet and PFNet for staged processing, and SLSR and BGNet for joint training.Source: §2 · pp. 3–4
  • Five transformer decoding strategies differ in feature flow: progressive, dense, feedback, separate, and pyramidal.Feedback decoding combines lower-resolution features with outputs for global information, or can use higher-resolution feedback to focus on local features.Source: §2 · pp. 4–5, Figure 2 · p. 4
  • Prior transformer COD models apply varied decoding designs, including progressive fusion, iterative feedback, pyramidal shrinking, and separate high- and low-resolution processing.Examples include MSCAF-Net, HitNet, FSPNet, and TPRNet.Source: §2 · p. 5
  • Transformer-based COD methods are categorized by main approach and decoding strategy in Table 1.The related-work discussion also describes transformer use across multi-stage, multi-scale aggregation, and joint-training approaches.Source: §2 · pp. 5–6, Table 1 · p. 5

3 Methodology

MSRNet extracts features at multiple input scales, selectively integrates them, and recursively refines them in a decoder. Its components are designed to preserve scale-specific information, strengthen contextual learning, and improve camouflaged-object predictions.

Multi-Scale Feature Encoder

  • PVTv2 extracts features at 1×, 1.5×, and 2× input scales, with larger auxiliary scales intended to help detect small and tiny objects.The encoder produces four feature maps per scale, one for each encoder stage.Source: §3 · p. 7

Scale Merging Network

  • ABSIUs align and merge features at each resolution, using multi-head spatial attention to emphasize relevant scales and regions.The four attention heads generate scale-specific maps, weight their corresponding features, and sum the processed features.Source: §3 · pp. 7–8

Recursive-Feedback Feature Refinement Decoder

  • The recursive-feedback decoder combines progressive decoding with feedback from lower-resolution maps to subsequent higher-resolution maps, preserving global context.Progressive decoding combines feature maps from lowest to highest resolution.Source: §3 · p. 9, Figure 6 · p. 9

Multi-Granularity Fusion Unit

  • MGFUs refine decoder features by integrating multi-granularity inputs through grouped processing and cross-channel interactions.The final MGFU output passes through an upsampling and convolutional COD head, whose sigmoid output represents pixel-level object probabilities.Source: §3 · pp. 9–10, Figure 7 · p. 11

Loss Function

  • Training combines binary cross-entropy with Uncertainty Awareness Loss, weighted by λ, which increases gradually using a cosine strategy.The uncertainty term is described as penalizing predictions with high uncertainty.Source: §3 · pp. 10–11

4 Experiments

Experiments use four COD datasets and compare MSRNet with state-of-the-art methods. The model reaches state-of-the-art results on two datasets, ranks second on the other two, and is visually compared on challenging object sizes and scenes.

4.1 Experiment Setup

  • Evaluation covered CAMO, CHAMELEON, COD10K, and NC4K, totaling 10,513 COD images; 1,000 CAMO and 3,040 COD10K images were used for training, with the remainder reserved for testing.Source: §4.1 · p. 11

4.2 Results and Comparison

  • MSRNet achieves state-of-the-art results on CAMO10K and NC4K, and ranks second on CAMO and CHAMELEON; it typically has fewer parameters than models with the same backbone and configurations.The comparison covers static-image COD datasets and reports backbones, input dimensions, and parameter counts.Source: Table 2 · p. 12, §4.2 · p. 13
  • Figure 9 compares detections of multiple, small, and tiny objects; reported errors include minor false areas, missed small object parts, and overlooked fine details.The authors characterize these failures as minor and report better performance than competing methods on the same samples.Source: Figure 9 · p. 14, §4.2 · p. 14

5 Ablation Studies

The ablations compare decoder strategies, input shape, encoder, and input scales. Their results show performance differences across these component choices.

5 Ablation Studies

  • 5.12% performance increase for M9’s 1.0×, 1.5×, and 2.0× scales was the highest reported; higher scales support finer-feature detection.M8’s 1.0×, 1.5×, and 1.7× scales increased performance by 4.54%; closely spaced scales may provide similar features and less novel information.Source: Table 3 · p. 16, §5 · p. 16
  • 4.63% performance improvement for PVTv2-B4 (M6) exceeded PVTv2-B3’s 3.01% and PVTv2-B5’s 3.41%.Source: §5 · pp. 15–16, Table 3 · p. 16
  • 0.21% performance gain for recursive-feedback M1 contrasted with -0.73% for dense-progressive M2 and -0.35% for combined M3.The authors suggest unrefined multi-resolution aggregation may disrupt scale-specific features. B0 used PVTv2-B2, a 352×352 input, 0.5×/1.0×/1.5× scales, and progressive decoding; experiments changed one component at a time.Source: §5 · pp. 14–15, Table 3 · p. 16
  • 1.42% performance improvement for M4’s 384×384 input shape; larger inputs preserve fine details that may otherwise be lost.Source: §5 · p. 15, Table 3 · p. 16

6 Conclusion

MSRNet combines multi-scale feature extraction and recursive decoding to address small and multiple camouflaged objects. It achieves state-of-the-art results on two COD datasets and ranks second on two others, while requiring additional computation and remaining designed for static images.

6 Conclusion

  • Across four benchmark COD datasets and comparisons with 20 SOTA models, MSRNet achieved SOTA performance on two datasets and ranked second on the other two.Source: §6 · p. 16
  • MSRNet uses a Pyramid Vision Transformer for multi-scale features, Attention-Based Scale Integration Units for selective merging, Multi-Granularity Fusion Units, and recursive-feedback decoding.Source: §6 · p. 16
  • Multi-scale feature extraction requires additional computational resources, and the model is presently designed for static images; future work could target efficiency and video-based COD.Source: §6 · pp. 16–17
Loading 2511.12810v2…