|
|
Face superresolution from image sequence
Mezina, Anzhelika ; Rajnoha, Martin (oponent) ; Burget, Radim (vedoucí práce)
This work is focused on application of deep learning in increasing resolution of images containing face. This can be applied in different fields, including security. For example, in case of incident, the police needs to identify a culprit from the records of security camera. The aim of this work is to propose neural network models, which would work with sequence of frames, and to compare these models with existing methods for a single image super-resolution. For this purpose, a new dataset with sequences of the images with faces is created. The methods for the single super-resolution are trained on the new dataset. The new architectures for multiframe super-resolution are proposed. They are based on U-Net model. This model is successful for segmentation tasks, but it can be also applied for super-resolution tasks. To improve this architecture, the residual blocks and its modification are used. To avoid blurring effect and recover more details, the perceptual loss function is applied. In the first part of this work, the description of neural networks and overview of the architectures, which can be applied in super-resolution, is provided. The second part contains the methods for super-resolution of a single frame, multiframe, video. In the next section, there is a description of proposed architectures and description of the experiment. In the last part of the work, multiframe methods and single frame methods are compared. In the result, the proposed methods recover more details, however, some architectures produce artefacts, which can be reduced using a filter, for example, Gaussian. New methods allow to reduce the number of failed face recognition. This fact is necessary for person identification in case of incidents.
|
|
|
Face superresolution from image sequence
Mezina, Anzhelika ; Rajnoha, Martin (oponent) ; Burget, Radim (vedoucí práce)
This work is focused on application of deep learning in increasing resolution of images containing face. This can be applied in different fields, including security. For example, in case of incident, the police needs to identify a culprit from the records of security camera. The aim of this work is to propose neural network models, which would work with sequence of frames, and to compare these models with existing methods for a single image super-resolution. For this purpose, a new dataset with sequences of the images with faces is created. The methods for the single super-resolution are trained on the new dataset. The new architectures for multiframe super-resolution are proposed. They are based on U-Net model. This model is successful for segmentation tasks, but it can be also applied for super-resolution tasks. To improve this architecture, the residual blocks and its modification are used. To avoid blurring effect and recover more details, the perceptual loss function is applied. In the first part of this work, the description of neural networks and overview of the architectures, which can be applied in super-resolution, is provided. The second part contains the methods for super-resolution of a single frame, multiframe, video. In the next section, there is a description of proposed architectures and description of the experiment. In the last part of the work, multiframe methods and single frame methods are compared. In the result, the proposed methods recover more details, however, some architectures produce artefacts, which can be reduced using a filter, for example, Gaussian. New methods allow to reduce the number of failed face recognition. This fact is necessary for person identification in case of incidents.
|
host ::
RSS.