IMAGE STYLE TRANSFER BASED ON A PRE-TRAINED VGG16 NETWORK: DETAILED FORMULATION, JUSTIFICATION FOR THE CHOICE OF NETWORK AND QUANTITATIVE EVALUATION

Neural style transfer involves representing the content of one image in the visual style of another. This paper presents a comprehensive and reproducible formulation of style transfer via optimisation, based on a pre-trained VGG16 network, in line with the work of Gatys et al. The content is represented by the activations of the relu4-2 layer and the style by the normalised Gram matrices of five layers (relu1_1 to relu5-1). The generated image is initialised using the content image and optimised using the Adam optimiser. We justify the choice of VGG16 based on its architecture and a controlled comparison with VGG19, ResNet-50 and the fast AdaIN method across 16 content–style pairs. Evaluation uses SSIM, PSNR, style distances (Gram matrices and colours) and computation time. Across the 16 pairs, the method preserves the content structure (average SSIM of 0.319) whilst transferring the style, and converges within a few hundred iterations. VGG16 matches VGG19 in terms of quality at a lower cost and significantly outperforms ResNet-50; it preserves the structure better than AdaIN without any training, albeit at the cost of increased computation time. All parameters are specified and the code is provided.