Criminal Identification using Residual Densenet with Particle Swarm Optimization for Facial Attribute Prediction

Abstract: Criminal identification using machine learning has
gained a lot of attention for how it can be used to enhance
security and individualization. Facial analysis in particular is
offering meaningful information about age, gender and
emotional condition, important information for effective
identification systems. Traditional deep learning models are
difficult to be applied in real-time applications due to their high
computational complexity and low flexibility to dynamic
datasets. There is a need for models which are not only
accurate, but also efficient and generalizable to diverse
conditions. This paper proposes a new method combining
Residual DenseNet and Particle Swarm Optimization (PSO) in
order to improve the accuracy and the compute speed of facial
attribute prediction. Residual DenseNet is used as the backbone
for the deep feature extraction whereas Particle Swarm
Optimization (PSO) is used for the optimization of
hyperparameters and model architecture for capturing fine
granular facial details. The model is evaluated on three datasets
including FaceForensics++, DrawSomething and Flickr-Faces-
HQ which represents a wide range of expressions, poses and
lighting conditions. The proposed system exhibits great
improvements in the way that the system can predict the age,
gender and emotional states as opposed to normal based
models. It also shows better generalization between data sets
ensuring robustness under a real-world situation. Visual
examination in form of feature map and quantitative
evaluation in form of accuracy, F1, precision and recall values
validate the model.
Keywords: Criminal identification,