Table 1.

A summary of previous studies on face-PAD with comparison with our proposed method.

Category	Detection Method	Strength	Weakness
Uses still images	-Uses handcrafted image features [15,16,17,18,19,21,22,23,24,25]	-Detection system is simple and easy to implement -Can achieve high processing speed	-Detection performance is limited because of handcrafted features designed by humans based on limited observation aspects of face-PAD problem
	-Uses deep image features: CNN [20,26]	-Uses deep features extracted by CNN for enhancing detection performance	-More complex and requires more power and processing time than the methods that only use handcrafted image features
	-Uses a combination of deep and handcrafted image features [27]	-Uses a very deep CNN network to efficiently extract image features -Uses SVM for classification instead of fully-connected layer that might reduce the overfitting problem. -Higher detection performance using a combination of deep and handcrafted image features	-More complex and requires more power and processing time than the methods that only use handcrafted image features
Uses sequence images	-Uses stacked CNN-RNN network to learn the temporal relation between image frames for face-PAD [28]	-Obtains higher detection performance than previous methods that only use a still image for detection using information learnt from more than one image	-Complex structure requiring more power and processing time - The CNN network is shallow with only two convolution layers and one fully connected layer
	-Uses very deep stacked CNN-RNN to learn the temporal relation between image frames -Combines deep and handcrafted image features to enhance the detection performance (Proposed method)	-Uses very deep CNN network to efficiently extract image features for inputs of RNN -Obtains higher detection performance than previous methods using very deep CNN-RNN network and handcrafted image features	-Requires more power and processing time to process a sequence of images