CIAFill: Lightweight and Fast Image Inpainting with Channel Independent Attention

Image inpainting is a classic technique in computer vision research. The quality of image inpainting has improved significantly since the advent of convolutional neural networks. However, this approach generally results in blurry and semantically inconsistent reconstruction because of operating valid and invalid pixels with equal weight. The gated convolution computing feature attention is proposed to resolve this issue but this attention mechanism was less efficient and computationally expensive. This study proposed CIAFill that alleviates this problem using channel independent attention. This mechanism applied channel attention to each channel for activating valid channels and reduced the computational cost to the dimension of the channel. The proposed architecture included a channel attention generator and a channel attention projection PatchGAN that utilize the channel independent attention mechanism. This study proved that CIAFill could successfully train the inpainting model with 1/1600 smaller gating parameters than the earlier gated convolution-based study. CIAFill achieved comparable performance to other feature attention-based approaches in the experiments on CelebA-HQ and Places2 datasets.