Enhancing object clarity in single channel night vision images using deep reinforcement learning
Robbani, Mohammad Elham
Sazid, Md. Riaz Ul Haque
Siam, Sk. Shahiduzzaman
MetadataShow full item record
There are a lot of novel approaches to image processing using Machine learning and classical image processing. But most of them take a huge dataset[like machine learning] or they are slow and ineﬃcient[like plain image processing]. Keeping this in context it was always a center of attraction to solve the problem of denoising and clarity enhancing in night vision images. As night vision images are like an asset sometimes, most of the CCTV footage is considered in this. Because if we think simply a complete footage of a CCTV has 50 percent of the recording in day time and 50 percent in night time. And just like that there is 50 percent chance of capturing any event at night. Now if we consider a crime scene which has to be extracted from the cctv footage at night time there is a huge probability of the footage to contain noise , distortion and clarity compromisation. In these scenarios identity extraction is diﬃcult. But along the progress of computation we have image processing and machine learning to develop and ﬁlter these images but we talked about the disadvantages before. To train a big dataset for extracting 1 footage is not feasible. So, we are considering a new novel approach which is also considered as state-of-the-art approach of using AI to ﬁlter a noisy night time single channel image and enhancing clarity and retain identity in it. In this work we will be facing limited resources and gradually developing those images by training an intelligent agent based on reward bias.So that after training the agent a limited resource it can predict future pixels based on it’s reward bias. Our approach on processing the images will consider deep Q learning and using a convolutional network based on Q learning. Our Aim will be to retain most of the information dealing with these limited resources using a reinforcement learning based approach built with the above stated structure.