The 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. The proposed LightFuse model consists of two sub-networks: a GlobalNet and a DetailNet. GlobalNet is to learn global information on the spatial dimension, whereas DetailNet aims to enhance local details on the channel dimension. Both GlobalNet and DetailNet are based solely on depthwise convolution and pointwise convolution to reduce required parameters and computations. LightFuse is trained with two extreme exposure LDR images to avoid problems such as large storage requirements, processing time, and power budget caused by a sequence of LDR images.
https://www.hackster.io/dmitrywat/machine-learning-inference-on-raspberry-pico-2040-e6e874 Trying to download TFLite Micro Need to do this on a linux machine as opposed to windows.
- 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
python train.py
@article{liu2021lightfuse,
title={LightFuse: Lightweight CNN based Dual-exposure Fusion},
author={Liu, Ziyi and Yang, Jie and Yanushkevich, Svetlana and Yadid-Pecht, Orly},
journal={arXiv preprint arXiv:2107.02299},
year={2021}
}