Enhancing Negative Film Colorization through Systematic CycleGAN Architectural Modifications: A Comprehensive Analysis of Generator and Discriminator Performance

This research addresses the urgent need for deep learning-based negative film colorization technology through systematic modifications to the CycleGAN architecture. Unlike conventional approaches that focus on colorizing black-and- white images, this study targets the conversion of digitized negative film images, which present unique challenges such as color inversion and detail restoration. The dataset consists of 500 negative images (train A), 500 unpaired color images (train B), as well as 5 negative images and 5 color images for testing purposes. The entire dataset was obtained from personal scanning efforts. 19 architectural modifications were proposed and tested individually, without simultaneously implementing all changes. The primary focus was on developing network structures, without utilizing external evaluation metrics such as SSIM, PSNR, or FID. Modifications included the addition of residual blocks, alterations in filter quantities, activation functions, and inter-layer connections. The Evaluation was conducted qualitatively and based on generator and discriminator loss values. The most optimal modification (Modification 4) demonstrated significant loss reduction (G: 2.39–4.07, F: 2.82– 3.66; D_X: 0.36–0.93, D_Y: 0.15–1.39), yielding more accurate and aesthetically pleasing color images compared to the baseline architecture. The fundamental cycle consistency loss structure was maintained to ensure the unpaired training capability remained intact. This research demonstrates that careful architectural modifications can significantly enhance negative colorization results, while simultaneously creating opportunities for the future development of deep learning-based digital image restoration technologies.