Reflection Removal using Single Images

Abstract

Image optimization has continued to be the highlight of the visual computing world. It is revolutionizing the visual world as we see it. With the world becoming more and more digitized, images have become the basic source of information for human activity. In image processing, the quality of an image is quite often compromised, leveraged by several external elements. This may be due to the poor lighting conditions or the interference of some transparent medium or any other natural phenomenon like rain, fog, etc. Among all these ill-posed problems, reflection removal has also caught the interest of a massive audience of the scientific community. The presence of reflections in images or videos results in undesired alterations and quite a loss of information. It not only is unpleasant to the eyes but also a hindrance to many of the computer vision tasks. From detection to classification, from recognition to localization to tracking, removal of reflections is of utmost importance. All things considered; we need to devise a solution which not only serves the purpose of enhancing the quality of images but also helps in pre-processing of the images for computer vision tasks. Various approaches have been proposed for reflection removal; from specialized hardware to computational techniques that further branches from conventional methods, exploiting physical properties of reflection, refraction and such, to deep learning models. Different aspects of previous approaches have been explored to achieve a framework that demonstrates the implementation in terms of increased quality. This work focuses on the minimization of the reflection of single images using conventional methods. As this remains a challenging task to-date, with a huge scope of performance improvement. We have proposed a method that depends upon the observation, that there exists an inconsistent blurring between the background and the reflection layers. This assumption has been the core of many reflection removal methods because of its visual significance for removal techniques. Further elaborating, the proposed algorithm adopts two-stage thresholding. The first stage deals with the uniform thresholding and the second stage is cruder in its form. Continuous thresholding gives a washed-out effect in the output image. To counter that, TV regularization approach decomposes the input image into the texture and structural part. To implement the input image’s texture to the output image, the textural part is redefined using a mask. Mask is obtained through DCT filtering where the scene details are preserved. To better align the texture layer of input and structure layer of output, a soft matting technique is applied. The final result of the texture layer is enhanced using Dark channel prior and then simply added to the structural layer to get the results. Evaluation between the prevailing state-of-the-art techniques was conducted both based on the visual analysis and quantitatively. For experimentation, images prepared by Rose Lab at Nanyang Technological University, Singapore, were utilized. It includes a large range of diverse, input images contaminated with reflections and ground truth for both reflection and background layer. Both real-world and synthetic images are included in the dataset. Some of the images tested have also been downloaded from the internet. Quality metrics against which the comparison with the other algorithms have been made are PSNR (Peak Signal-to-Noise Ratio), MSE (Mean Squared Error), and SSIM (Structural Similarity Index). Other than these full reference quality measures some of the non reference quality measures are also calculated which includes BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator), NIQE (Naturalness Image Quality Evaluator), and PIQE (Perception based Image Quality Evaluator). Output results are quite promising.