StGAN: A Novel Symbolic Signal Decomposition Base on GANs and Swin Transformer
Ming Sun, Li Guo, Long Chen · 2023
The symbolic imagery signal decomposition is a common problem in digital signal processing. Its main purpose is to divide the symbolic imagery signal into different parts. However, in real-world applications, symbolic imagery signal captured by the camera is usually influenced by complex negative lighting environments such as highlights, reflections, drop shadows, or information loss. To overcome those problems, a generative adversarial network with Swin Transformer (StGAN) is proposed and applied to symbolic imagery signal decomposition and semantic segmentation tasks. In addition, a realistic image dataset taken in complex lighting conditions is proposed for symbolic imagery signal decomposition, which have complex lighting environments and edge information loss, we name it the MLT dataset. We demonstrate StGAN brings significant improvements in performance than some existing methods in the accuracy of symbolic imagery signal decomposition on MLT datasets. Further experiments on Sky and Facades datasets prove that StGAN works well in other tasks, especially when fewer data is involved in training, but it can still ensure good accuracy.