Images of road scene in low-illumination environment are often loss the details revealing the essence of images which increases the danger when driving at night, especially in those areas without any light. Therefore, an efficient deblurring, enhancement algorithm for low-illumination images is necessary. Here is a CNN based model to restore the low-illumination image. Following pictures show some enhancement examples from human-generated dark image, and the real night img of bdd100k and eurocity dataset
The architecture of our proposed light enhancement network (LE-net) was inspired by many advanced technologies, including linear bottleneck and inversed residual("mobilenetV2"), self attention distillation("attention distillation"), and feature pyramid module("FPN") were used to improve the network performance. During training, we randomly add gaussian blur, gaussian noise to enhance the generalization performance of the model.
tensorflow >= 1.6.0
imgaug ==0.2.8
opencv-python >= 3.3.1.11
scikit-image >=0.15.0
Please download the corresponding dataset from the official, and extract the daytime images (both training and test dir) in a specific directory, and then set the directory in dataset/bdd_daytime.py
Below script gives you an example of training a model with our models.
python train.py --batch_size=10 --learning_rate=1e-3 --f_log_step=20 --f_summary_step=20 --f_save_step=2000
Below script gives you an example of visualing the results after training a model.
CHECKPOINT_DIR_NAME: something like /usr/my_dir/checkpoint
python predict.py --checkpoint_dir=CHECKPOINT_DIR_NAME
Below script gives you an example to evaluate the model after training.
CHECKPOINT_DIR_NAME: something like /usr/my_dir/checkpoint
python eval.py --checkpoint_dir=CHECKPOINT_DIR_NAME