A Novel Approach Towards Photorealistic Image Stylization

• Published in: International Conference on Computing, Communication, and Networking Technologies (Online) pp. 1 - 7
• Main Authors: Mishra, Anish, Ghosh, Bidita, Shekhar, Shashank, Baliarsingh, Santos Kumar
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.07.2023
• Subjects: Convergence; Distortion; Image stylization; Measurement; Neural networks; Optimization; Photorealism; Visualization
• ISSN: 2473-7674
• DOI: 10.1109/ICCCNT56998.2023.10306667

Photorealistic image stylization - a technique which involves transferring the artistic style from one image onto another, while simultaneously preserving the inherent content of the target image. Numerous methods for achieving photorealistic image stylization have been developed, however they produce stylizations that are spatially inconsistent with noticeable artifacts and distortion in output images which are not present in captured photographs. We propose novel advancements to address these problems and enhance photorealism to improve the existing methods by incorporating both content and style loss in the optimization process. We propose two key enhancements: activation function enhancement and learning rate optimization. In the activation function enhancement,the conventional ReLU activation function is replaced with the Exponential Linear Unit (ELU) activation function. Furthermore, we conduct a comprehensive learning rate optimization study to customize the learning rate for our method. Both enhancements lead to improved convergence speed and solution quality, resulting in superior photorealistic images. The dataset used in this research paper is the Deep Photo Style Transfer(DPST) dataset which contains a collection of content and style images that can be used to train neural networks to perform photo style transfer and the evaluation of this model was done using several metrics, including perceptual quality, visual similarity, and structural similarity. Through rigorous evaluation and comparison with existing methods, our enhanced approach achieves remarkable and visually appealing stylization results that are more preferred by human subjects. Our novel approach achieves a higher SSIM score (the higher the better) as discussed in Table 1 than other existing methods. It also opens up new possibilities for integrating desired artistic styles into images while preserving their original content.