A Symmetric Dual-Attention Generative Adversarial Network with Channel and Spatial Features Fusion

Abstract

Many existing generative adversarial networks (GANs) lack effective semantic modeling, leading to unnatural local details and blurring in generated images. In this work, based on DivCo, we propose a Symmetric Dual-Attention Generative Adversarial Network (DivCo-SDAGAN) with channel and spatial feature fusion in which the Dual-Attention Module (DAM) is introduced to strengthen the feature representation ability of the network to synthesize photo-realistic images with more natural local details. The Channel Weighted Aggregation Module (CWAM) and the Spatial Attention Module (SAM) of the DAM are designed to capture the semantic information of channel dimension and spatial dimension, respectively, and they can be easily integrated into other GANs-based models. Extensive experiments show that the proposed DivCo-SDAGAN can produce more diverse images under the same input, achieving more satisfactory results than other existing methods.