Research on application icons generation based on GAN improved by self-attention mechanism

Abstract

Aims The software interface design is changing and updating and has spawned a large number of interface icon design requirements. Using deep learning and generative adversarial network(GAN) to assist icon design can simplify the process of icon design and enrich the creativity of icon conception. Methods The Python tool was used to capture nearly 2000 icon data and perform a simple filtering. After the steps of icon preprocessing, data transformation, structure design of generating network and discriminant network, definition of loss function, model recording and saving, a generation adversarial network suitable for icon design is successfully constructed. Results In this study, icon generation is well explored by combining deep learning and generative adversarial networks. Although the change of function loss and the output icon quality are not ideal in the preliminary practice, higher quality and new icons are obtained by improving the icon recognition algorithm, and the function loss value is reduced. The research has certain guiding significance for more personalized, higher quality and efficient icon generation and design in the future. Conclusion It is feasible to apply deep learning and GAN into icon design. The newly generated icons have certain artistic expression and instructional intention, and can be used as ideas and creativity in icon design process.