Üretken Çekişmeli Ağ Kullanarak Eski Görüntüleri Renklendirme ve İyileştirme

Abstract

– Nowadays, data for deep learning is crucial to the success rate of the study results. With the improvements we can make to image processing for this need, it is intended to achieve better quality and complete images and to translate black and white images into a color image, which may be one of the problems we face with older data, and to complete missing pixels. Thanks to the user interface developed in the study, it serves both as a user's ability to edit personal images and as a supporting tool for other studies. As a method, pre-processing, deep learning and training and retouching are used in order. One of the varieties of Generative Adversarial Network (GAN) architecture, the Pix2pix model has been developed as a way to transform from source image to target image. From this point of view, the Pix2pix model has been the ideal network for improvement and conversions of images. The highest performance result in experiments was achieved for indoor and outdoor classification with accuracy of 82%. In addition, the highest Structural Similarity Index Measure (SSIM) value was 0.9256 and the lowest Peak Signal to Noise Ratio (PSNR) value was 65.11 dB when the average values were taken into account with the experiments to improve indoor colorization, outdoor colorization, human face colorization, image repair and image quality. Our study includes discussions and assessments based on scientific findings.