BIECON

A blind image evaluator based on a convolutional neural network (BIECON) is a no-reference image quality assessment method using a CNN. This code implements the system described in the following paper:

J. Kim and S. Lee, “Fully deep blind image quality predictor,” IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 1, pp. 206–220, Feb. 2017.

Prerequisites

This code was developed and tested with Theano 0.9, CUDA 8.0, and Windows.

Generating local quality score maps

Set BASE_PATH to the actual root path of each database. Set FR_MET_BASEPATH and FR_MET_SUBPATH in gen_local_metric_scores.m.
For each database, data will be stored in "FR_MET_BASEPATH + FR_MET_SUBPATH".
Run gen_local_metric_scores.m using Matlab. We provide a SSIM metric as default.

Environment setting

Setting database path:

For each database, set BASE_PATH to the actual root path of each database in the following files: IQA_BIECON_release/data_load/LIVE.py, IQA_BIECON_release/data_load/TID2008.py, and IQA_BIECON_release/data_load/TID2013.py.

(These BASE_PATH should be same to the BASE_PATH in gen_local_metric_scores.m.)

Setting local quality score map path:

Set FR_MET_BASEPATH and FR_MET_SUBPATH_{DB name} in IQA_BIECON_release/data_load/data_loader_IQA.py. {DB name} can be LIVE, TID2008, or TID2013

(These should be same to those in gen_local_metric_scores.m.)

Detailed configuration of local quality score maps is set in NR_biecon.yaml.

fr_met: This describes the name of the full-reference image quality assessment metric. The corresponding local quality score maps must be generated first. ex) SSIM, FSIM ...
fr_met_scale: This indicates the scale ratio of the local quality score maps to their original images. This should be same to image_resize in gen_local_metric_scores.m.
fr_met_avg: If True, for each divided image patch, local quality score maps are averaged to be scalar values. Otherwise, patch of local quality score maps are used. This is used to reduce the usage of memory.

Training BIECON

We provide the demo code for training a BIECON model.

python example.py

tr_te_file: Store the randomly divided (training and testing) reference image indices in this file.
snap_path: This indicates the path to store snapshot files

Quantitative results

BIECON was tested on the full-sets of LIVE IQA, TID2013, and CSIQ databases. During the experiment, we randomly divided the reference images into two subsets, 80% for training and 20% for testing. The correlation coefﬁcients were averaged after the procedure was repeated 10 times while dividing the training and testing sets randomly.

Database	SRCC	PLCC
LIVE	0.9603	0.9622
TID2013	0.7205	0.7650
CSIQ	0.8251	0.8380

daimuuc / IQA_BIECON_release