🌍 Cross Branch Feature Fusion Decoder for Consistency Regularization-based Semi-Supervised Change Detection

Welcome to the official repository of our paper "Cross Branch Feature Fusion Decoder for Consistency Regularization-based Semi-Supervised Change Detection"! Our paper has been accepted by IEEE ICASSP 2024.

Our method introduces CBFF (Cross Branch Feature Fusion) — a novel decoder that synergistically combines the local expressiveness of CNNs and the global reasoning of Transformers, enabling high-performance change detection even under extreme label scarcity. Through comprehensive experiments on WHU-CD and LEVIR-CD, we demonstrate consistent superiority over seven state-of-the-art semi-supervised methods.

---

🔍 Introduction

Semi-supervised change detection (SSCD) aims to detect pixel-level changes between bi-temporal remote sensing images using a small amount of labeled data and abundant unlabeled imagery — a critical setting given the high cost of manual annotation.

While recent works explore Transformer-based architectures for global context modeling, we observe (as shown in Fig. 1 of our paper) that convolutional decoders significantly outperform transformer-only decoders under low-label regimes (e.g., 5% labeled data). This is because Transformers require large volumes of annotated samples to converge stably.

To bridge this gap, we propose CBFF: a hierarchical decoder that fuses two complementary branches:

• Local Convolutional Branch (LCB): Learns robust local features efficiently with minimal supervision.
• Global Transformer Branch (GTB): Captures long-range contextual dependencies across image pairs.

By combining these branches via residual fusion and enforcing strong-to-weak consistency regularization on unlabeled data, CBFF achieves superior performance without requiring massive labeled datasets.

---

🏗️ Framework

Our network follows a structured pipeline:

• Difference Feature Generator: Uses a Siamese ResNet backbone to extract features from two input images; computes change features via absolute difference and two convolution layers.
• Bottleneck (ASPP): Applies Atrous Spatial Pyramid Pooling on the deepest feature to capture multi-scale context.
• CBFF Modules (×3): Hierarchical fusion blocks integrating LCB and GTB at each resolution level, enabling stable learning with few labels.
• Dual Prediction Heads: Two parallel classifiers generate change maps from fused features, enabling consistency loss during training.

Left: CBFF decoder fuses CNN and Transformer features at each level.
Right: Training uses supervised CE loss on labeled data and unsupervised consistency loss on unlabeled data.

---

📊 Results

1. Quantitative Results on WHU-CD and LEVIR-CD

The two numbers in each cell denote the changed-class IoU and overall accuracy (OA), respectively.
Results are reported under semi-supervised settings with varying labeled ratios: 5%, 10%, 20%, 40%.
Bold values indicate state-of-the-art performance.

🔹 WHU-CD

Method	5%	10%	20%	40%
AdvEnt	57.7 / 97.87	60.5 / 97.79	69.5 / 98.50	76.0 / 98.91
s4GAN	57.3 / 97.94	58.0 / 97.81	67.0 / 98.41	74.3 / 98.85
SemiCDNet	56.2 / 97.78	60.3 / 98.02	69.1 / 98.47	70.5 / 98.59
SemiCD	65.8 / 98.37	68.0 / 98.45	74.6 / 98.83	78.0 / 99.01
RC-CD	57.7 / 97.94	65.4 / 98.45	74.3 / 98.89	77.6 / 99.02
SemiPTCD	74.1 / 98.85	74.2 / 98.86	76.9 / 98.95	80.8 / 99.17
UniMatch	78.7 / 99.11	79.6 / 99.11	81.2 / 99.18	83.7 / 99.29
Ours	81.0 / 99.20	81.1 / 99.18	83.6 / 99.29	86.5 / 99.43

🔹 LEVIR-CD

Method	5%	10%	20%	40%
AdvEnt	67.1 / 98.15	70.8 / 98.38	74.3 / 98.59	75.9 / 98.67
s4GAN	66.6 / 98.16	72.2 / 98.48	75.1 / 98.63	76.2 / 98.68
SemiCDNet	67.4 / 98.11	71.5 / 98.42	74.9 / 98.58	75.5 / 98.63
SemiCD	74.2 / 98.59	77.1 / 98.74	77.9 / 98.79	79.0 / 98.84
RC-CD	67.9 / 98.09	72.3 / 98.40	75.6 / 98.60	77.2 / 98.70
SemiPTCD	71.2 / 98.39	75.9 / 98.65	76.6 / 98.65	77.2 / 98.74
UniMatch	82.1 / 99.03	82.8 / 99.07	82.9 / 99.07	83.0 / 99.08
Ours	82.6 / 99.05	83.2 / 99.08	83.2 / 99.09	83.9 / 99.12

✅ Our method consistently outperforms all baselines, achieving +1.5%~+2.8% IoU gain on WHU-CD and +0.3%~+0.9% IoU gain on LEVIR-CD over UniMatch.

---

2. Visualization Results

Qualitative comparison on selected samples from WHU-CD and LEVIR-CD (5% labeled training ratio):

Detection results of different methods under 5% labeled ratios.

---

🚀 Getting Started

Installation

cd CBFF
conda create -n pytorch12 python=3.10.4
conda activate pytorch12
pip install -r requirements.txt
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 -f https://download.pytorch.org/whl/torch_stable.html

Pretrained Backbone

Download the pre-trained ResNet-50 checkpoint: ResNet-50.

├── ./pretrained
    ├── resnet50.pth

Dataset

• WHU-CD: imageA, imageB, and label
• LEVIR-CD: imageA, imageB, and label

Please modify your dataset path in configuration files.

├── [Your WHU-CD/LEVIR-CD Path]
    ├── A
    ├── B
    └── label

✅ Update the dataset path in your config file (e.g., config.yaml) before training.

---

🛠️ Usage

▶️ Training & Testing

cd CBFF
bash scripts/train.sh
bash scripts/test.sh

Modify parameters in scripts/train.sh:

• dataset: Set to whu or levir.
• split: Set to 5%, 10%, 20%, 40%.
• method: Use supervised for supervised training, fixmatch_CbffDecoder for consistency training.

---

📝 Citation

If you find this project useful, please consider citing:

@inproceedings{xing2024cross,
  title={Cross branch feature fusion decoder for consistency regularization-based semi-supervised change detection},
  author={Xing, Yan and Xu, Qi’ao and Zeng, Jingcheng and Huang, Rui and Gao, Sihua and Xu, Weifeng and Zhang, Yuxiang and Fan, Wei},
  booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={9341--9345},
  year={2024},
  organization={IEEE}
}

❤️ Acknowledgements

This project is based on SemiCD and UniMatch. Thank you very much for their outstanding work.