Generative synthesis of logos across DCT domain

Abstract While GAN-based generative logo synthesis has achieved great success, their fundamental operating principle in generating logos in pixel domain suffers from the limitations of spatial scope for both learning and optimization, especially in terms of exploiting those pixels located far away from the centre in question. Generative learning in pixel domain has achieved great success in exploiting their correlations in processing images towards desired objectives, yet learning in frequency domain could provide added benefits in exploiting pixel correlations without worrying about their spatial locations and increasing their modeling costs. In this paper, we analyze the spectral bias from a frequency perspective to overcome such limitations and hence propose a dynamic self-adaptive optimization on GAN-based generative learning, leading to a dynamic and generative logo synthesis in DCT domain. To achieve exploitation of all the pixel correlations inside the whole image regardless of their spatial locations, we introduce an approximated DCT transformation and decompose both the input images and the generated images into relatively independent DCT frequency bands. As a result, a new channel of DCT domain generative learning can be established to support the existing pixel domain learning towards improved logo synthesis. Since learning across different frequency band constantly varies, we further propose a dynamic optimization scheme to maximize the effectiveness of contributions from each individual DCT frequency band. Extensive experiments are carried out and the results in comparison with the existing state of the arts illustrate that our proposed achieves significant superiority in terms of both synthesized logo quality, integrity and variety.