|
Portrait matting is a challenging task in the field of computer vision, which has important applications in meeting background replacement, image editing and other scenarios. However, how to obtain high precision alpha matte, especially for the part of the edge transition area, has been a difficult problem in this field. Focusing on the post processing part of the encoder-decoder, this paper designs a new efficient patch refinement network(PRN). It first recieves a coarse alpha matte and repair the selected patches under the guidance of flaw map, these patches will be reassembled to origin alpha matte to obtain high precision results. At the same time, this paper introduces ConvGRU into the refinement layer, which can improve the temporal consistency when migrating the refinement network to the task of video portrait matting. Experiments show that the PRN allows the original image to be sent to the network after down-sampling, which can improve the performance of the network inference while recovering the high-resolution details of the alpha matte as much as possible. |
|
Keywords:Artificial Intelligence; Deep Learning; Portrait Matting; Edge Refinement; Temporal Consistency |
|