Face Image Synthesis

Abstract

 Facial image synthesis has made rapid dynamic
progress with the fast expansion of deep learning techniques.
Reference samples can give complete primary information about
texture and content in this job and improve the visual quality of
synthetic images. A normalized network with a multi-scale
pyramid structure is used in this instance. The dual-channel
normalization architecture at the center of the normalization
network is capable of obtaining previous knowledge about
various semantics from reference samples. There are two
conditional normalization branches in the DNC specifically.
Through the first branch, the reference image can be spatially
adaptively normalized based on the input image's semantic mask.
The second branch is used to normalize the adaptive
representation of the modified input image on the reference
image. By dividing the complete cross-domain mapping into two
branches, DNC may highlight the distinct significance of
structural and spatial elements. To avoid information
redundancy and improve the final performance, the Gated
Channel Attention Fusion module is used to differentiate and
merge useful information from the two branches. This generated
synthetic image is then compared with photos of criminals in the
crime database. If a match is found, the image will be displayed
along with details about that criminal. Comparing the images is
done by using pillow library with image hashing