Share The Best: Fast Photo Selection and Enhancement System

By Yilei Cao, Yixing Gao, Nora Horanyi, and Hyung Jin Chang

Abstract

Automatic photo selection has been an actively studied area in computer vision research during the last decade. This task is very laborious for humans, especially when the photo album is very large. No-reference image aesthetic quality assessment (IAQA) networks are designed to solve this task by automatically assessing a given image based on aesthetic metrics. Despite good performance, most existing methods rely on deep networks and require plenty of computing resources when dealing with a large number of photos. In this work, we combined a state-of-the-art IAQA network and an image retouching network to build a fast image selection and enhancement system. To reduce the computing resources, we light-weighted the IAQA network and improved its pre-processing method. In addition, by exploring the relationship between IAQA networks and image retouching networks, we built the image retouching network with fast running speed and good enhancement performance. This image selection and enhancement system uses limited computing resources and runs fast on a non-GPU device.

Contribution

We present a fast image selection and enhancement system that can process a large number of photos with limited computing resources.
We trained and compared image aesthetic quality assessment networks based on different lightweight CNN models. We presented an image preprocessing method that is suitable for images with arbitrary aspect ratios.
We explored the relationship between IAQA networks and image retouching networks. We proposed a new method to evaluate the performance of an image re-touching network using IAQA networks.

Framework

Demo Usage

Requirements

Python3, Requirements.txt

Build

3D-LUT trilinear package (More details)

cd trilinear_cpp
sh setup.sh

Run

cd PyQt_demo
python3 logic.py

IAQA Training

After downloading AVA benchmark, NIMA-train.ipynb can be run to train IAQA models based on PyTorch NIMA.

Trained model parameters can be found in model_parameter/para_iaqa

Image Enhancement

MIT-Adobe-fiveK dataset can be downloaded with image_retouching/download_5k.py after download txt file from MIT-5K

image_retouching/lut/lut_eval.py is to enhance the image folder with the 3D-LUT model

image_retouching/csrnet/csr_eval.py is to enhance the image folder with the CSRNet model

yileiCao/shareTheBest