PhD Student in Information Science at University of Colorado Boulder
A review of facial gender recognitionNg, CB., Tay, YH. & Goi, BM. A review of facial gender recognition. Pattern Anal Applic 18, 739–755 (2015). https://doi.org/10.1007/s10044-015-0499-6
Gender RecognitionThe authors posit gender as a significant attribute for applications of biometrics, surveillence, marketing, and HCI that would benefit from automated gender classification. In this paper, they discuss both social and technical challenges to automated gender classification (AGR) and present a survey on current methods and data used for AGR. In presenting this work, they discuss how humans can "easily" differentiate between "a male and a female," setting up the assumptions or worldview of gender in this piece. They also propose computer vision as non-intrusive, but not requiring human cooperation, as a benefit of computer vision.
- Human-computer interaction, in this case, personalization which the authors feel would be more "human-like" when recognizing gender.
- Video surveillence. They propose uses like restricting areas to single-gender and monitoring for rule breakers. They also seem to propose building gender segregation around monitoring systems.
- Biometrics, or basically face recognition augmented with gender. For example, if searching for a specific person, the database can be first assessed by gender if that person's gender is known.
- Demographic collection, particularly in the context of marketing. Basically collecting gender information for statistics.
- Advertisement and recommender systems, for recommending specific products by gender.
- Input image or video.
- Face detected. This process is decoupled from the gender classification algorithm, and thus the face detector performance will impact gender classification downstream.
- Image cropped to face / bounding boxed.
- Pre-processing steps as defined by researcher (e.g., normalization of contrast or brightness, cropping out regions like hair, geometric alignment, downsizing).
- Feature extraction.
- Binary classification (male/female)
Feature ExtractionThe authors posit this as the most important step and so dedicate a section to it. There are two approaches.
(1) Geometric: Based on measurements of facial landmarks (ficudial distances = the distance between points on the face).
(2) Appearance-based: "Based on some operation or transformation performed on the pixels of an image. This can be done at the global or local level. At the global level, features are computed from the whole image resulting in a single feature vector. In local feature extraction, the image is partitioned beforehand into some arbitrary regions (which may be equally spaced or otherwise) or into semantically meaningful regions such as eyes, nose and mouth areas." The authors state global approaches are best for when gender features are not known prior. Local approaches focus on regions of the face (e.g., periocular). Appearance-based approahces include: pixel intensity values, rectangle features, local binary patterns, scale invariant feature transform (SIFT), gabor wavelets, and a few others named in the paper. They also discuss using external cues, like hair and clothing, as an option.
ChallengesTechnical: Head pose, lighting, image quality.
Social: Age, ethnicity, expression, occlusion, facial hair ("But since these can be different between the genders, they may also provide useful discriminative cues.")