An Improved Image Inpainting Method Based on Feature Similarity Context Encoder
Zhiqiang Li, Zhouzhong Zhang, Hongchen Guo · Journal of Physics Conference Series · 2018
With the development of deep learning techniques, image inpainting is a hot and popular field to use information around the residual images for repairing the missing area. Context encoder (CE) makes great progress in this field. However, it will cause the issue that the model tends to be overfitting when repairing the image. The existing method can't address it very well. In this paper, we propose a feature similarity context encoder (FSCE), which integrates the feature extraction network with the encoder. FSCE uses feature loss to improve the performance of features and avoid overfitting problem. Various experiments between FSCE and CE are conducted on VOC 2007 and Paris Street View datasets. The results demonstrate that our FSCE model achieves superior performance in both the texture of output and dealing with the overfitting issue. FSCE model outperforms CE in the aspect of PSNR and SSIM.