ngshya / low-quality-image-detection

Low Quality Image Detection using Machine Learning

Low Quality Image Detection

This repository contains some experimental Python code designed for the detection of low-quality images through a machine learning approach. In particular, the algorithm will perform

• sharp (good quality) images detection,
• defocused blur detection,
• motion blur detection,
• horizontal/vertical bands detection,
• noise detection,
• excessive exposure detection,
• glare detection,
• dark photo detection,
• uninformative constant colour detection.

Check the Medium article for the technical details.

Environment setup

Create and use a Python environment with Python 3.10.13 and the packages listed in requirements.txt. Check here or, if you are using Anaconda, here for the detailed steps.

Dataset

The images are available here and here. Once downloaded the .zip folder, extract its contents ( defocused_blurred, motion_blurred, sharp) inside a folder called dataset_dms (unless another location is chosen).

Synthetic data generation

python 01_synthetic_low_quality_images.py

A folder called dataset_synthetic (unless another location is chosen) will be created with generated low quality images.

Features generation

python 02_features_generation.py

A CSV file called df_public.csv (unless another name is used) will be generated.

Models training

Look at 03_models_{name}.ipynb notebooks as examples.

Just give me the script!

python detect.py {image path}