A COMPARATIVE EVALUATION OF CNN ARCHITECTURES FOR SINGLE-IMAGE GANS

Authors : Eyyüp Yıldız, Erkan Yüksel

Pages : 194-205

Doi:10.62301/usmtd.1736275

View : 39 | Download : 140

Publication Date : 2025-12-26

Article Type : Research Paper

Abstract :Generative Adversarial Networks (GANs) have achieved remarkable success in image synthesis, enabling the generation of photorealistic and diverse visual content. While most generative models depend on large datasets to capture visual variability, single-image GANs such as SinGAN demonstrate that rich generative behavior can emerge from the internal patch statistics of a single natural image. However, the effect of convolutional neural network (CNN) backbones on single-image generative performance remains underexplored. This study presents a comparative analysis of five CNN architectures—Inception, ResNet, DenseNet, CBAM, and MobileNet—integrated into the SinGAN framework to investigate their influence on image quality, diversity, and computational efficiency. Each architecture was trained under identical multi-scale SinGAN settings using 15 natural images, and evaluated with Single-Image Fréchet Inception Distance (SIFID), Multi-Scale Structural Similarity (MS-SSIM), and Learned Perceptual Image Patch Similarity (LPIPS), complemented by qualitative visual assessment. The results reveal consistent trade-offs among backbones: ResNet best preserves global structural coherence; DenseNet maximizes fine-detail diversity through dense feature reuse; CBAM enhances perceptual realism via attention module; Inception provides balanced multi-scale feature representation; and MobileNet achieves strong diversity and quality with favorable computational efficiency. These findings demonstrate that architectural design fundamentally governs generative behavior in single-image GANs. The study provides empirical insights and practical guidelines for selecting CNN backbones based on trade-offs between quality, diversity, and efficiency-supporting the design of data-efficient generative models for real-world and resource-constrained applications.
Keywords : Tek görüntü GAN, SinGAN, CNN mimarileri, görüntü sentezi, SIFID, MS-SSIM, LPIPS

ORIGINAL ARTICLE URL