Transforming design elements and enhancing visual communication via machine learning-based graphic design

This paper proposes convolutional efficient transformer-based image feature extraction. First, it combines convolution from different angles to introduce translation invariance, scale invariance, and locality. Then, a lightweight convolution module with depth-wise convolution and atrous convolution is used to change the traditional processing of the input image of the transformer to accelerate the convergence speed and improve the stability. Aiming at the problem that cycle Generative Adversarial Network (CycleGAN) has low texture clarity when processing images, the convolutional efficient transformer-based image feature extraction algorithm is added to the generator of GAN to enhance the effect of CycleGAN in image style transfer. Results demonstrate that the proposed algorithm performs well regarding top-1 accuracy, number of parameters, floating-point operations and convergence rate. It can effectively analyse and classify visual elements, providing valuable insights and aiding in tasks such as image categorisation, style transfer, generative design or automated design recommendation systems.