Table 1.

Summary of previous studies on body-image-based gender recognition.

Categories	Methods	Strength	Weakness
Using a pre-designed (hand-designed) feature extractor for extracting image features.	-Using gait or 3D shape information [13,14,15,16].	-High recognition accuracy can be obtained.	-Requires a series of human body images. -Requires the cooperation of users in the image acquisition step. -Uses a predesigned feature extraction method (features). -Requires an expensive capturing device (scanner) to obtain 3D information of the human body [15,16].
	-Using the HOG or BIFs feature extraction method in only single visible-light images [10,11].	-Easy to implement.	-Limits recognition accuracy because of the use of predesigned and weak feature extraction methods (HOG and BIFs features).
	-Using HOG feature in the combined visible-light and thermal images [5].	-Easy to implement. -Enhances recognition accuracy by utilizing both visible-light and thermal images of the human body.
	-Using a weighted HOG feature in combined visible-light and thermal images [12].	-Compensates the effects of background regions on recognition accuracy by applying weight values on HOG features. -Enhances recognition accuracy by utilizing both visible-light and thermal images of the human body.	-Limits recognition accuracy because of the use of a predesigned and weak feature extraction method (weighted HOG feature).
Using a leaning-based feature extractor method for extracting image features (proposed method)	-Learns the feature extractor using CNN for extracting image features.	-Extracts the more suitable image features for recognition using a pre-trained feature extractor model based on CNN. -Higher recognition accuracy can be obtained compared to predesigned feature extractor methods, such as HOG, BIFs, or weighted HOG.	-Needs training time to train the feature extractor (CNN model).