Image Generation Method of Bird Text Based on Improved StackGAN
Li Hui, Yuan Xuchang · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
Text to image is a difficult task, which needs to focus on the semantic relationship between text and image. On account of this, an improved StackGAN network is applied to generate bird images according to text in study. Based on StackGAN twostage network, the network is optimized by deep separable convolution to reduce the amount of model parameters; SE attention mechanism is embedded in the model to enhance the key features of the image and suppress the invalid features; The residual structure is used to deepen the network and improve the generation effect. The experimental results show that the average IS of the algorithm on the CUB dataset is 3.94 and the FID is 36.47. Compared with StackGAN, the IS is increased by about 5%, the FID decreases by about 30%, and the model size is only about 1/2 of StackGAN, which improves the reasoning speed and generation effect.