Single Image De-raining Using Convolutional Neural Network

Abstract

Rain streaks vary in size, quantity, and direction, making removing them from individualimages difficult. Recent advancements in deep learning, especially thoseusing CNN-based techniques, have shown promising results in addressing this issue.However, the requirement for additional consideration of the rain streaks locationinformation in the image is a significant drawback of these methods. Methodsbased on deep learning have proven to be quite effective in handling syntheticand real-world rainy images. These methods use convolutional neural networks(CNNs) to their full potential to learn the correspondence between rainy and rainfreeimages. We typically use an encoder-decoder architecture where the encoderpulls features from the rainy image and then creates the rain-free image using thelearned features. These algorithms can efficiently learn the complicated correlationsbetween rain streaks and ground truths by training on large-scale datasetsthat combine images with and without rain. End-to-end methods aim to train asingle model that converts the rainy image into its rain-free counterpart withoutexplicitly decomposing it into the rain and the background components. Additionally,researching end-to-end approaches offers a fascinating way of improvingthe de-raining algorithm�s efficiency. More effective and efficient techniques forremoving rain streaks from single images will probably be developed when thisresearch study continues to be investigated.