Enhancing colour and brush stroke pattern recognition for stable abstract art generation using modified deep convolutional GANs.

This article focuses on the development and evaluation of a modified Deep Convolutional Generative Adversarial Network (mDCGAN) architecture for generating high-quality abstract art images, emphasizing brushstroke and color pattern analysis. The proposed mDCGAN introduces architectural adjustments to improve training stability and reduce common issues like mode collapse and gradient vanishing, outperforming the standard DCGAN in generating diverse and realistic artworks. The study also explores the latent space of generated images through random walks, revealing vector relationships between colors and brushstroke patterns, and conducts statistical analyses demonstrating significant differences in output stability before and after 250 training epochs. These findings suggest the potential of mDCGAN to enhance digital art generation, with future work aimed at employing larger GAN architectures and advanced pattern recognition techniques to further improve artistic output quality.