ARNet-v2

April 6, 2026 · View on GitHub

Assisted Refinement Network Based on Channel Information Interaction for Camouflaged and Salient Object Detection

Authors:

Camouflaged Object Detection (COD) remains a challenging task in computer vision, aiming to identify and segment objects that are visually integrated with their surroundings. Although existing methods have achieved progress in cross-layer feature fusion, two key issues persist during decoding: (1) insufficient cross-channel information interaction within same-layer features, which limits feature expressiveness, and (2) ineffective collaborative modeling between boundary and region information, leading to incomplete regions and inaccurate boundaries.

To address these challenges, we propose the Assisted Refinement Network (ARNet) — a dual-dimensional collaborative decoding framework that enhances both feature interaction and boundary–region consistency.

Key Contributions

Channel Information Interaction Module (CIIM) Introduces a bidirectional horizontal–vertical integration mechanism at the channel dimension, enabling feature reorganization and interaction across channels to capture complementary cross-channel information effectively.

Boundary Extraction (BE) and Region Extraction (RE) Modules Generate boundary priors and object localization maps, respectively, and employ Hybrid Guided Attention (HGA) within CIIM to jointly calibrate decoded features, thereby improving boundary sharpness and spatial localization.

Multi-scale Enhancement (MSE) Module Enriches contextual feature representations using a three-branch multi-scale convolution strategy, expanding the receptive field and improving global–local feature integration.

Experimental

Extensive experiments on four COD benchmark datasets demonstrate ARNet’s state-of-the-art performance and strong generalization ability. We further transfer ARNet to the Salient Object Detection (SOD) task and validate its adaptability across various downstream applications, including polyp segmentation, transparent object detection, and industrial/road defect detection.

📖 Table of Contents

🎯 Our Motivation

image

📝 Overview

image

💻 Environment

  • python = 3.9
  • Other packages can be found in requirements.txt.

🛠️ Materials & Evaluation

Required Materials

Evaluation Script

We provide a simple script to reproduce the quantitative results.

📂 File Structure

Please organize your files as follows to run the script without any modifications:

.
├── eval
├── evaluation
│   └── evaluation.py
├── TestDataset
│   └── NC4K
│       └── GT
│           ├── <mask_1>.png
│           ├── <mask_2>.png
│           └── ...
└── test_maps
    └── ARNet
        └── NC4K
            ├── <pred_1>.png
            ├── <pred_2>.png
            └── ...
  • TestDataset/NC4K/GT/: Contains the ground truth masks for the NC4K dataset.
  • test_maps/ARNet/NC4K/: Contains the predicted maps from our method (ARNet) on the NC4K dataset.

🚀 How to Run

Simply execute the evaluation.py script to automatically compute and display all evaluation metrics.

python evaluation.py

Note: You can also modify the mask_root and pred_root variables within the script to evaluate different datasets or methods.

📊 Experimental Results

image