Constructing Ghost Free High Dynamic Range Images Using Convolutional Neural Network and Structural Similarity Index

A foreign object, commonly called as a ghost artifact, is integrated in the HDR output image when there is a moving object in the photography scene. The problem is persisting even after numerous models proposed by researchers. The most advanced techniques for capturing HDR photographs are still struggling with the produced output till date. The majority of the existing techniques are unable to tackle the situation where luminance varies in the input images. This ultimately causes the algorithms to compromise on quality by reducing the input images. In this research, two techniques are presented that help in getting rid of ghost artifacts from HDR output. The first method that is used in this research is a vintage method that uses structural similarity index measurement. The second method is using Convolutional Neural Network that is proved to be the best configuration of neural network for image recognition purpose. The first method using structural similarity is based on the vintage method of matching two images. This method works by comparing the objective image with the reference image and the algorithm estimates the degree of similarity in context of luminance, contrast and structural information. The second method is the use of convolutional neural network. Since convolutional neural network is a specialized tool for image processing, a convolutional neural network is designed and trained to estimate the index of similarity of two or more images.