الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 102 |  |  |
| | | from | 102 | | |
Abstract
Image colorization is defined as the problem of assigning colors for grayscale images. Due to the success of applying Deep Learning techniques in different applications, it is used to solve the colorization problem. Adopting Deep Learning in image colorization proved to be a promising approach that might show future breakthroughs.
Image colorization solutions can be classified according to several criteria as input image type, number of colored output images, colorization methodology, techniques or networks used in colorization and network paths. However, most of the solutions are classified in literature according to one or two criteria.
The thesis presents a solid review that targets all these criteria to classify the solutions simultaneously with a considerably large number of papers. Additionally, the review shows the most commonly used measuring metrics of comparison as well as the used datasets. The findings of the review reveal that Deep Learning has become a widely used approach in solving the colorization problem.
Despite the large number of papers that solved the fully automatic colorization problem, many of them failed to accurately colorize images with several objects. The reason behind this could be dealing with the multi-objects in the image as a single whole image, regardless the variety of objects’ colors it contains. This thesis is considered as an extension to the efforts exerted in the last few years to start the colorization solution with an object detection phase. Along with this, colorization of each individual object will be considered as well as the colorization of the full image. After colorizing the objects and the full image, they are fused together to attain higher colorization accuracy. In our work, Scaled-Yolov4-P7 detector was used due to the superiority of its accuracy to the state-of-the-art detector, which might have positive effects on the colorization quality. On comparing the results of our model to literature, it is found that adopting Scaled-Yolov4-P7 increased the PSNR by 2.6%. Also, the results of colorized images with different extensions are compared, and png extension got 5.8% better value of Learned Perceptual Image Patch Similarity (LPIPS) metric than JPEG.