How to perform low quality image detection (for instance, blur detection, glare detection or noise detection) using machine learning and deep learning.
Low-quality image detection is an interesting machine learning problem because it addresses real-world challenges across diverse applications (for instance, blurry image detection in surveillance systems or automatic quality check while taking photos with a smartphone). The quality of images can significantly impact the outcomes of various downstream tasks, making the development of effective detection algorithms crucial.
In this tutorial we try to build a machine learning model able to detect whether a photo has any quality issue.
Image quality issues may include: blurriness, presence of bands, noises, over exposure, glare, darkness, etc.
Whenever we are trying to perform a blur detection, a glare detection or a noise detection, we can suppose that all bad quality photos of the same kind should share same common properties. The traditional image processing approach consists in building and applying filters and measures to detect these common properties. These approaches are stable, fast, work on most of the cases, but they are based on one single metric. Just to be clear, I’m not saying that traditional approaches are less valid than machine learning ones. Instead, I strongly believe that it depends on the context and the data. Here, we just want to experiment with a machine learning based approach. The code base of this tutorial is available on GitHub.
There are not many public datasets available for our problem setting: dms dataset (public domain license), blur detection dataset, and images with quality flaws dataset. In this tutorial, we will use the first dataset. It contains…