Automatic Detection of Random Holes for Image Inpainting based on Generative Adversarial Networks

Abstract

Image inpainting is mainly used to repair the missing parts of the image or remove unwanted objects in the image In the past the common inpainting method was to repair or reconstruct the corrupted area content through the texture and line structure in the image Inpainting makes the whole image reach the human visual and can't tell if it has been repaired In the past few years deep learning methods have made significant progress in image restoration In this study an end-to-end model established by the generative adversarial networks was proposed for image inpainting It is different from the generative adversarial networksarchitecture used in image inpainting issues as others Usually the higher quality inapinting model were based on corrupted image and the mask of corrupted area However in some situations the corrupted area image is not easy to obtain Hence the architecture proposed in this study was increased the function to detect mask automatically The proposed method also improves the quality of the result By automatically detecting the mask the corrupted areas are completed in accordance with the image semantics as a result user does not need to perform additional pre-processing actions on the corrupted area when using the model Moreover compared with other architectures the architecture proposed in this study does not reduce the patching performance due to the deepening of the network layer The experimental results show that the small corrupted area inpainting has the same excellent effect as other studies; but in a large and continuous corrupted area the results have higher quality performance than others In addition the model can also be used to implement image outpainting

Date of Award	2019
Original language	English
Supervisor	Ming-Shi Wang (Supervisor)