Deep image inpainting
Over the years, many techniques have emerged to reconstruct and modify images for a myriad of applications. One ingenious application is image inpainting, which is to restore the missing parts of an image. The latest approach employs deep learning technique to solve the problem. Deep convolutional n...
Saved in:
| 主要作者: | |
|---|---|
| 其他作者: | |
| 格式: | Final Year Project |
| 語言: | English |
| 出版: |
2019
|
| 主題: | |
| 在線閱讀: | http://hdl.handle.net/10356/79001 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 機構: | Nanyang Technological University |
| 語言: | English |
| 總結: | Over the years, many techniques have emerged to reconstruct and modify images for a myriad of applications. One ingenious application is image inpainting, which is to restore the missing parts of an image. The latest approach employs deep learning technique to solve the problem. Deep convolutional neural networks are used to capture the abstract details of many training images, so that it can guess the context of a missing region. The performance of the network heavily relies on the information provided upon training. Most work failed to utilize or realize the importance of prior information that may boost the proficiency of neural networks. This project attempts to use segmentation maps as a feature engineering to create supplementary information to aid the image inpainting process. The method of inpainting process proposed will consist of two stages. First is to generate the segmentation maps of the missing region. Second is to take the prior segmentation maps generated for fusion into the inpainting process. Training and evaluation are done on ADE20K dataset with eight categories of segmentation defined and all other bodies as the background category. |
|---|