Image Pre-Processing Detection Using Deep Reinforcement Learning

Reinforcement learning is a popular domain of Machine learning being actively used in many industries. With proper reward function and problem formulation, a task can easily be solved using Reinforcement learning.  Computer vision domain often face the problem of noisy data. Even when the model is trained on good quality data, during inference, the noisy data can create problem and thus causing model to fail. An essential step to overcome the problem of bad data is pre-processing of data, but pre-processing of image data is itself a complex problem and requires subject matter experts to decide which kind of preprocessing should be applied on a given image for a particular compute vision task.  To solve this problem of choosing correct pre-processing for a given images, we have proposed a novel approach of automation of image pre-processing using deep reinforcement learning.  This approach is developed and tested for one of the most popular problems in image data, which is noise in images, while it has also shown potential how it can change the whole scenario of machine learning models being applied in the field of computer vision.