Publication: Identity-of-interest (ioi) recognition from image sketches in video surveillance system
Loading...
Date
2022-07-01
Authors
Lim, Mu Khai
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Human face is a unique ID everyone has since birth. Generally, everyone has the same face orientation, which are: having two eyes, having one nose, and having one mouth. Despite this, the small difference in face features make everyone has a unique face. With this property, face recognition system can gain face recognition ability after learning each of the face features, providing a highly effective and accurate method in recognizing the identities of every person. Moreover, face recognition can be implemented in assisting the law enforcement where the police can search for suspect’s photo from their database quickly and effectively as well as identity tracking on suspects through video surveillance
system. However, in most cases, the photo image of a suspect is not available, therefore face recognition can only be performed on sketch drawing based on the recollection of
eyewitnesses to identify the suspect. Retrieving a corresponding image from forensic sketch remains a great challenge even with state-of-the-art model available today due to
the modality difference between sketch and photo is high. Thus, in this project, a sketch-based face recognition surveillance system is developed where the system is able to
identify the suspect based on suspect’s sketch image. The system is fed with suspect’s sketch image and suspect’s footage. The sketch image is used to train the Inception-ResNet face recognition model. The suspect’s body will first be segmented out from the video using DeepLabv3+. Then face detection is performed on the segmented image and face is
cropped out. Photo-to-sketch synthesis model from CycleGAN is used to convert the modality of the cropped face to pseudo-sketch image. The pseudo sketch image is used to
perform face matching with the sketch image mentioned earlier. As a result, the finalizedsystem has recorded an identification accuracy of 81.1% on CUFS dataset. However, when
frontal video of personal data is used, the accuracy is only about 60.2%. This indicates that it still remains a challenge facial recognition system to accurately recognize a person
from just a sketch photo.