Extended multi-scale feature fusion and balanced generative adversarial network for image inpainting under limited data

Image inpainting has made remarkable progress in recent years due to the availability of sufficient training data. However, when the training data is insufficient, i.e., image inpainting in few-shot regimes, the performance of existing image inpainting methods degrades drastically. Two challenges impede image inpainting under few-shot regimes, adequate information and unstable training. To address the above issues, this paper proposes a new method for image inpainting under limited data. Specifically, an Extended Multi-scale Feature Fusion (EMFF) module is proposed to capture representative contextual information by progressively reducing the number of channels for features with increasing receptive fields, thereby utilizing effective information. Additionally, we leverage a multi-scale discriminator whose crosslayer features are mixed for supplementation to formulate a Balanced Generative Adversarial Network (BGAN) to stabilize the training dynamics. Comprehensive experiments demonstrate the effectiveness of the proposed method.