DAS-Net: A Lightweight Dynamic Convolution Network with Attention Gates and Deep Supervision for UAV Semantic Segmentation

Young Jae Kim and Sang-Chul Kim

Kookmin University

Overview

DAS-Net extends ThinDyUNet with three architectural improvements for UAV semantic segmentation:

1. Symmetric Dynamic Convolution — DyConvBlock in both encoder and decoder
2. Attention Gates — filter skip connections to suppress irrelevant features
3. Deep Supervision — auxiliary loss heads at last 3 decoder stages (λ=0.4)

Results

Test Set Performance (20K training samples, full test set ~168K images)

Model	Params (M)	Precision	Recall	Dice	mIoU	ms	FPS
ThinDyUNet	1.34	0.9597	0.5937	0.6407	0.5731	8.72	114.7
PSPNet	21.4	0.8553	0.6686	0.6768	0.6094	13.80	72.5
PAN	21.4	0.9055	0.7047	0.7045	0.6438	14.77	67.7
DeepSupDyUNet	1.34	0.9364	0.6838	0.7084	0.6485	8.67	115.3
FullDyUNet	1.63	0.8794	0.7301	0.7394	0.6706	9.49	105.4
UNet	24.4	0.9003	0.7333	0.7413	0.6760	14.16	70.6
DAS-Net (Ours)	1.66	0.8408	0.7700	0.7506	0.6786	9.41	106.3

Ablation Study

Model	Sym. Decoder	Attn Gate	Deep Sup	mIoU	Improvement
ThinDyUNet (baseline)	—	—	—	0.5731	—
FullDyUNet	✓	—	—	0.6706	+17.0%
DeepSupDyUNet	—	—	✓	0.6485	+13.2%
DAS-Net (Ours)	✓	✓	✓	0.6786	+18.4%

Dataset

We use the UAV semantic segmentation dataset proposed by Kim and Jang (Appl. Sci. 2025).

• 605,045 paired visible light and infrared images
• Binary segmentation masks (UAV vs. background)
• Training: 20,000 randomly sampled (seed=42)
• Validation: 1,000
• Test: full 168,143 images

Model Architecture

Component	Details
Encoder	7-stage DyConvBlock + MaxPool
Decoder	7-stage DyConvBlock + Bilinear Upsample
Attention Gate	At each skip connection
Deep Supervision	Last 3 decoder stages (λ=0.4)
Channels	64 (fixed)
DyConvBlock	K=2 kernels, τ=30, GroupNorm(8) + LeakyReLU
Parameters	1.66M

Requirements

torch
torchvision
torchinfo
omegaconf
segmentation_models_pytorch
einops
tqdm

How to Run

Training

python train_dasnet.py          # DAS-Net (proposed)
python train_thindyunet.py      # ThinDyUNet (baseline)
python train_fulldyunet.py      # FullDyUNet (ablation)
python train_deepsupdyunet.py   # DeepSupDyUNet (ablation)
python train_unet.py            # UNet
python train_pan.py             # PAN
python train_pspnet.py          # PSPNet

Testing

python test_dasnet.py
python test_thindyunet.py
python test_fulldyunet.py
python test_deepsupdyunet.py
python test_unet.py
python test_pan.py
python test_pspnet.py

Configuration

Each model has a corresponding config file (config_*.yaml) with dataset path, hyperparameters, and checkpoint settings.

Training Details

• Input: 512×512
• Optimizer: AdamW, LR: 1e-4
• Loss: DiceLoss
• Batch size: 24
• Epochs: 50 (early stopping, patience=30)
• Scheduler: ReduceLROnPlateau (factor=0.5, patience=15)
• GPU: NVIDIA A6000
• Threshold: 0.5

Citation

@article{kim2025dasnet,
  title={DAS-Net: A Lightweight Dynamic Convolution Network with Attention Gates and Deep Supervision for UAV Semantic Segmentation},
  author={Kim, Young Jae and Kim, Sang-Chul},
  journal={Applied Sciences},
  year={2025}
}

License

MIT License