Transfer Learning Using Convolutional Neural Networks for Face Anti-spoofing

A Database for Face Presentation Attack Using Wax Figure Faces

“…Similar to the results in Table 4, the ResNet-50 based PAD method performed the best, achieving 21.83% ACER, while the reflectance properties based [20] and VGG-16 based [27] method have no advantages in distinguishing real faces from wax figure faces. The overall results of the seven evaluated PAD methods on the WFFD database are shown in Table 6.…”

“…Shape based 3D mask PAD methods use shape descriptors [19,36,15] or 3D reconstruction [38] to extract discriminative features from faces and 3D masks. Instead of extracting hand-crafted features, deep feature based methods [29,27,34,23] automatically extracts features from the face images, and trend to have a higher detection accuracy and a better generalization ability.…”

“…Several face PAD methods were evaluated on the WFFD database to show how they can work for the proposed super-realistic 3D presentation attacks. These PAD methods were based on different features, including the multi-scale LBP [11], the color LBP [10], the Haralick features [1], the reflectance properties [20], the multi-level Local Phase Quantization (LPQ) [7], deep features based on the ResNet-50 model [37] and VGG-16 model [27]. They all achieved high detection performance against 2D spoofing attacks or 3D mask attacks.…”

Prescribing challenges following bariatric surgery

“…And the popularity of the transfer learning paradigm for pre-trained CNNs has pushed performance boundaries in various areas. Some researchers have introduce this paradigm into the PAD area and obtained promising results [15]. Also, the effectiveness of LSTM for modelling time-related information has been demonstrated in various areas.…”

“…The temporal changes are compressed and represented by a single image using the Motion History Image algorithm [19]. The texture patterns inn the spatio-temporal domain are modelled by using Local Binary Patterns (LBP) [11] and Convolutional Neural Networks (CNN) [15]. The proposed work offers a new direction to produce PAD-related information by efficiently encapsulating the distinct spatio-temporal information to texture patterns and significantly decreasing the computational effort required for modelling dynamic textures.…”