Abstract:
Face photo-sketch matching is an important
problem for law enforcement agencies in terms of
identifying suspects. In this study, a new sketch-photo
generation and recognition technique is proposed by
using residual convolutional neural network
architecture. The suggested RCNN architecture
consists of 6 convolutions, 6 ReLU, 4 poolings, 2
deconvolution layers. The proposed architecture is
trained with face photos and sketches. Sketches are
supplied as an input to the RCNN architecture and,
generated face photos are obtained as the output.
Then, the generated face photos are compared with
the photos of the people in the database. Structural
Similarity Index (SSIM) is used to measure the
pairwise similarity and the photo with the highest
index score is matched. CUHK Face Sketch Database
containing 188 images is tested. In the experiments,
148, 20, and 20 images are used for training,
validation, and testing, respectively. Data
augmentation applied to 148 training images
produced 444 images. Experimental results show that
the success of the training curve is 90.55% and the
validation success is 91.1%. True face recognition
success from generated face images with SSIM is
93.89% for CUHK Face Sketch database (CUFS) and
84.55% AR database.
