This code implements a model to predict linear motion blur parameters using a regression Convolutional Neural Network.
The Requirements.txt file is based off a conda environment. To create a conda environment based on Requirements.txt
conda create --name --file requiremnts.txt
This folder has the scripts to create the blur datasets. The dataset is based off COCO 2014
COCO_blur_augmentation_TestVal and COCO_blur_augmentation_Train create the Test/Val and Train datasets to train the model. The code randomizes COCO images which means it wont create the exact blurred images we used. The code is parallelized to use multiple cores. The number of images you get in train will depend on the number of cores you use. Each core yields approximately 21K images. The code will also export a csv file with the labeled data for filenames, angle, and length blur parameters.
COCO_blur_using_csv.py reads in one of the csv files to create the exact blur images used in the paper available at https://arxiv.org/abs/2308.01381. This is a slower method of creating images which will not be optimal for generating training datasets, but can be optimal to test with the same subset datasets as the paper available at https://arxiv.org/abs/2308.01381.
VGG16_Regressor.py Trains the regressor model.
parent_dir variable needs to point to the directory where all your blur datasets are located.
train_dir, val_dir, test_dir are the variables that concatenated the parent_dir with the specified folder for each dataset.
train_labels, val_labels, test_labels are the csv file names for each dataset that will be used by the data generator to create the batches with its labeled data.
Results.py Will run the results of your trained model, and calculate the weights_filenames for the weights you enter.
The deconvolution method was implemented from Zoran using this EPLL code