Daytime to Night Image Translation Based on Semantic Segmentation

Image-to-image translation is the process of converting images from one domain to another. The goal of image-to-image translation is to learn the mapping between input and output images. In the face of complex scenes, for example, when the field changes greatly from day to night, the daytime image scenes have complex semantic information, which makes it difficult for convolutional neural networks to learn specific texture details, resulting in ambiguities and other phenomena after image translation. Semantic segmentation is a visual scene understanding task. It understands and recognizes the image content from the pixel level, and then performs image segmentation according to the semantic information. The purpose is to predict the category label of each pixel in the input picture, so as to provide clearer and structured semantic information. In view of the mature development of semantic segmentation technology, we propose a day-to-night image translation method based on semantic segmentation. Through unsupervised learning, semantic information is provided by images and semantic segmentation graphs to generate more reasonable nighttime domain images.

I. Introduction

When humans observe an image through the visual system, they can integrate high-level semantic information from the image, convert the information into knowledge, and use knowledge to deal with more complex tasks. For computers, the method of deep learning can be adopted to learn high-level semantic information of pictures with large amounts of data, and then complex tasks can be processed with the help of high-level semantic information. Image translation is an important subject of computer vision, image translation of two input images, to achieve the generated image has one of the content of the other style. Image translation has a wide range of applications, not limited to art or graphic design. For example, converting a shot image into another image to create composite data is useful for training autonomous driving models. In map design, the model is capable of performing two transformations, satellite view to map and inverse mapping. Image inversion can also be applied to architecture, where models can make suggestions on how to complete unfinished projects.