PyTorch implementation for image quality assessment in the LIVE dataset. It is based on the Neural Image Assessment (NIMA) by Talebi and Milanfar.
As mentioned before the LIVE dataset is used composed of 1162 images. Download the data and convert its labels to a .txt file with the following structure:
imgName, imgMOS, imgStd
This will be later used for generating a the necessary vectors for the groundtruth.
As defined in NIMA, the Earth Mover's Disctance (EMD) is used as a ranked loss function. EMD receives two cumulative distribution functions to estimate the error. For this project the output of the network is a vector of probabilities of size 10, so the groundtruth needs to be a vector of probabilities of size 10. To do this, the .txt file with the MOS will serve the purpose to generate a truncated normal distribution using the imgMOS and imgStd in the range 1-10.
Additionaly, Mean Squared Error could be used as loss function with the same probability vector.
MSE and EMD could be combined for a more robust error function. These terms could be combined through the weighting of each term. See line 80 of main.py for further understanding.
To train a model, the script main.py should be used. In the scripts, there is a dictionary of parameters that could be changed including architecture to use, number of epochs, batch size and folder list.