PROJECT TITLE :

Single Image Reflection Removal Using Convolutional Neural Networks

ABSTRACT:

Specular reflection occurs when individuals photograph through glass, obscuring the view behind the glass. Most investigations have focused on recovering the transmitted scene from many photos rather than a single image because of the ease of implementation. It's not practicable for most people to use numerous photographs in real life because of the key shooting conditions. Single-image reflection reduction is proposed using convolutional neural networks in this paper. As a result of our ghosting model, acquired photographs have a reflective quality to them. Ghosting and relative intensity are used to create multiple-reflection pictures from a single image. Encoder and decoder are then built into a complete end-to-end network We employ a combined training technique to learn layer separation information from the synthesised reflection pictures in order to optimise the network parameters. Both internal and external losses are taken into account when optimising the loss function. This network is then used to remove a single image's reflection. The proposed solution does not necessitate the use of handcrafted features or specular filters to remove reflections. For both synthetic and real photos, the suggested method is able to successfully eliminate reflections from both images, as well as achieve the top scores in peak SNR, structural similarity and feature similarities.