This research aims to develop a high-performance Neural Style Transfer (NST) system that artistically combines the structural information of a target image with the artistic patterns of a reference style image using deep learning. A convolutional neural network (CNN) architecture with Gram matrix-based feature correlation improves visual stylization in VG Gram. VG Gram collects local and global texture patterns, enabling it to replicate intricate artistic styles more accurately than previous models. The technique was tested using FLICKR8K pictures and a Kaggle Collection of Paintings from 50 artists with various content and style inputs. VG Gram accomplished quantifiable benchmarks such a Structural Similarity Index (SSIM) of 0.742, a Peak Signal-to-Noise Ratio (PSNR) of 23.8 dB, a Total Loss of 398.57 million, a Content Loss of 3.36 million, and a Style Loss of 395.21 VG Gram's fast inference time of 1773.78 ms and lack of floating-point operation (FLOPs) data made it ideal for CPU-based or low-resource edge environments where computational limitations limit operational feasibility. The system was conceived and implemented in Python utilizing TensorFlow and PyTorch for model implementation and training. Jupyter Lab was used for experimentation, visualisation, and performance evaluation, enabling flexible and interactive development and improvement.
Sheriff et al. (Mon,) studied this question.