Official codebase for LightFuse: Lightweight CNN based Dual-exposure Fusion. Contains demo (see demo.ipynb) and scripts to reproduce experiments.
To our best knowledge, this is the first lightweight HDR fusion algorithm that could be used in power and resource constrained edge-computing devices. It is challenging to train a lightweight model with fewer parameters and layers while maintain comparable performance. The proposed LightFuse model consists of two sub-networks: a CombiningNet
and a FilteringNet
. The goal of CombiningNet is to learn the channel-related information, whereas FilteringNet aims in combining the spatial information
. Both CombiningNet and FilteringNet is based on depthwise separable convolution
to reduce required parameters and computations. LightFuse is trained with extreme exposure images to avoid possible fail during inference phase.
- Python = 3.7.9
- TensorFlow = 1.15.0
- Opencv-python = 4.4.0.44
- Scipy = 1.5.2
- Matplotlib = 3.3.1
- Clone this repo:
git clone https://github.com/Taichi-Pink/LightFuse-Lightweight-CNN-based-Dual-exposure-Fusion.git
cd LightFuse-Lightweight-CNN-based-Dual-exposure-Fusion
python test.py
- Prepare TFRecord.
python FuDataset.py
- run train.py