A Study on High-Quality Style Transfer Using Edge-Enhanced Gaussian Smoothing Technique(EEGST) for Minimizing Perceptual Loss

Recent advancements in image generation technologies, particularly those involving style transfer, have garnered significant attention. Style transfer involves merging the characteristics of one stylistic image with the content of another to produce a novel image. Most image generation techniques, including style transfer, heavily rely on neural networks. However, the use of neural networks, particularly during processes such as up-sampling, can introduce artifacts. This paper proposes a method to mitigate artifacts in models utilizing neural networks, specifically focusing on the style transfer model proposed in "Perceptual Losses for Real-Time Style Transfer and SuperResolution" [5]. The mentioned model exhibits advantages in terms of speed but suffers from checkerboard artifacts or similar issues. The proposed approach effectively addresses these concerns, reducing artifacts while preserving the overall form of the image. The method involves combining Gaussian smoothing to handle artifacts introduced by neural networks while ensuring the preservation of edges. The simplicity and versatility of the proposed method make it applicable to various models. This paper presents a method to reduce artifacts in neural networkbased models, with a particular focus on the style transfer model benchmarked against "Perceptual Losses for Real-Time Style Transfer and Super-Resolution." The proposed approach proves effective in maintaining image structure while mitigating artifacts, showcasing its potential applicability across different models.