.png){kind=icon}
.png){kind=icon}
Return
MTCNN plus plus : A CNN-based face detection algorithm inspired by MTCNN
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-023-02822-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Increasing security concerns in crowd centric topologies have raised major interests in reliable face recognition systems globally. In this context, certain deep learning frameworks have been proposed till date, for example, Haar Cascade, MTCNN, Dlib to name a few. In this communication, we propose a deep neural network for reliable face recognition in high face density images. The proposed framework is inspired by multi-task cascaded convolutional neural Networks (MTCNN) and, hence the name MTCNN++. In this framework, we have modified the layer density with increasing the neuron count. All the three internal layers of MTCNN, viz. P-Net, R-Net, and O-Net layers and observe that the modified Net-Layer MTCNN (MTCNN++) perform equally well to the MTCNN library or better. Moreover, 20% dropout has been used for tuning the framework for better recognition of the faces, both in terms of face clarity and face count. MTCNN++ exhibits better results as the preprocessing is done dynamically in contrast to the previous versions. The training of the model was done on a dataset comprising of 113,586 human faces in a bucket of 9661 images. The comprehensive dataset comprised of photographs from varied events, thereby presenting multiple human expressions. The accuracy of the model varies from 87.7% (average of 12 faces per image) to 99.7% (average of 2 images per images). The proposed framework fares better with large face count per image. MTCNN++ has further been compared to other literary proposals, and the results are appreciable.
Keywords:
Real-time system
Face detection
Deep neural network
Computer vision
MTCNN
Haar cascade
Dlib
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K