ReviewFair comparison of skin detection approaches on publicly available datasets
Introduction
Skin texture and color are important signs that people use to understand variety of culture-related aspects about each other, as: health, ethnicity, age, beauty, wealth and so on. The presence of skin color in images and videos is a signal of the presence of humans in such media. Therefore, in the last two decades extensive research in the context of expert and intelligent systems has focused on skin detection in videos and images. Skin detection is the process of discriminating “skin” and “non-skin” regions in a digital image and consists in performing a binary classification of pixels and in executing a fine segmentation to define the boundaries of the skin regions. Currently, skin detection is a sophisticated process involving not only the training of models, but also numerous additional methods, including data pre-processing and post-processing.
Skin detection is used within many application domains: it is used as a preliminary step for face detection (Hsu, Abdel-Mottaleb, & Jain, 2002) and tracking (De-La-Torre, Granger, Radtke, Sabourin, & Gorodnichy, 2015), body tracking (Argyros & Lourakis, 2004), hand detection (Roy, Mohanty, & Sahay, 2017) and gesture recognition (Han, Award, Sutherland, & Wu, 2006), biometric authentication (i.e. palm print recognition) (Sang, Ma, & Huang, 2013), objectionable content filtering (Lee, Kuo, Chung, & Chen, 2007), medical imaging. In this work a comprehensive analysis is carried out of how different expert systems (including artificial intelligence, deep learning, and machine learning systems) are designed in order to deal with the skin detection problem.
A useful feature for the discrimination of skin and non-skin pixels is the pixel color; nevertheless, obtaining skin color consistency across variations in illumination, diverse ethnicity and different acquisition devices is a very challenging task. Moreover, skin detection, when used as preliminary step of other applications, is required to be computationally efficient, invariant against geometrics transformations, partial occlusions or changes of posture/facial expression, insensitive to complex or pseudo-skin background, robust against the quality of the acquisition device. The factor that worst influences skin detection is the color constancy problem (Khan, Hanbury, Stottinger, & Bais, 2012): i.e. the dependency of pixel intensity on both reflection and illumination which have a nonlinear and unpredictable behavior. To be effective when the illumination conditions vary rapidly, some skin detection approaches use image preprocessing strategies based on color constancy (i.e. a color correction method based on an estimate of the illuminant color) and/or dynamic adaptation techniques (i.e. the transformation of a skin-color model according to the changing illumination conditions). Static skin color approaches that rely on image preprocessing can only partially solve this problem and their performance strongly degrades in real-world applications. A possible solution is considering additional data acquired out the visual spectrum (i.e. infrared images (Kong, Heo, Abidi, Paik, & Abidi, 2005) or spectral imaging (Healey, Prasad, & Tromberg, 2003)), however the use of such sensors is not appropriate for all applications and requires higher acquisition costs which limit their use to specific problems.
Skin detection is a challenging problem and has been extensive studied from the research community. Despite the large number of methods, there are only few surveys in this topic: the works in (Kakumanu et al., 2007, Prema and Manimegalai, 2012) are quite old and cover only the methods proposed before 2005, the surveys in (Chen et al., 2016, Mahmoodi and Sayedi, 2016, Naji et al., 2018) are more recent and contain a good investigation of methods, benchmarking datasets and performance related to a period of about two decades. Anyway in none of the above surveys is there a fair comparison among methods using the same testing protocols and datasets. The aim of the present work is not limited to survey the most recent research in this field (which is now enriched of methods based on deep learning (Xu et al., 2015, Zuo et al., 2017, Kim et al., 2017a, Ma and Shih, 2018), but also, and above all, to propose a framework for a fair comparison among approaches.
In this research, a novel framework is proposed that integrates different skin color classification approaches and compare their performance and their combination on several publicly available datasets. The major contributions of this research work are:
- •
An exhaustive literature review of skin color detection approaches with a detailed description of methods freely available.
- •
The collection and examination of almost all the datasets with ground truth for skin detection available in the literature. Such collection, which includes more than 10,000 labelled images, can be useful to produce a training set for CNN models.
- •
A framework to evaluate and combine different skin detector approaches. The source code of the framework and many of the tested methods will be made freely available for future research and comparisons. The system can be tuned according to the target application: on the basis of the application requirement, the acceptance threshold can be tuned to prune a large percentage of false accepts at a small cost of reduction in genuine accepts or vice versa a larger number of false accepts can be admitted to maximize the number of genuine accepts. The framework includes training and testing protocols for most used benchmark datasets in this field.
- •
A fair comparison among the most recent research and methods in the skin detection field, using the same testing protocols, benchmark datasets and performance indicators. An evaluation of computation time of each method in order to perform a comparison also in terms of complexity. A discussion about performance can help researchers and practitioners in evaluating the approaches most suited to their requirements according to computational complexity, memory requirements, detection rate and sensitivity.
- •
Three different CNN architectures are trained for skin detection and the model are made available.
The arrangement of this paper is as follows. In Section 2 related works in skin detection are presented, including a discussion about taxonomy of existing approaches and a detailed description of the approaches tested in this work. In Section 3 the evaluation problem is treated, the most known datasets used for performance evaluation are listed and commented, testing protocols and performance indicators used in our experiment are discussed. In Section 4 the experiments conducted using the proposed framework are reported and discussed. Finally, Section 5 includes the conclusions and some future research directions.
Section snippets
Skin detection approaches
Several skin detection methods assume that skin color can be recognized from background colors according to some clustering rule in a specific color space. Even if this assumption can be valid in a constrained environment where both ethnicity of the people and background colors are known, it is a very challenging task in complex images captured under unconstrained conditions and when individuals show a large spectrum of human skin coloration (Kakumanu, Makrogiannis, & Bourbakis, 2007). There
Skin detection evaluation: Datasets and performance indicators
To assist research in the area of skin detection, there are some well-known color image datasets provided with ground truth. The use of a standard and representative benchmark is essential to execute a fair empirical evaluation of skin detection techniques.
A fair experimental comparison
A fair comparison among different approaches is very difficult due to the lack of a universal standard in evaluation: most of published works are tested on self-collected datasets which often are not available for further comparison; in many cases the testing protocol is not clearly explained, many datasets are not of high quality and the precision of the ground truth is questionable since sometimes lips, mouth, rings and bracelets have been labelled as skin. In this section, we carry out a
Conclusion and future research directions
In this work a new framework to evaluate and combine different skin detector approaches is presented and an extensive evaluation of several approaches is carried out on 10 different datasets including more than 10,000 labelled images. A survey of most recent existing approaches is carried out, three well-known deep learning models for data segmentation are trained and tested to this classification problem and four new ensembles based on the combination of nine methods (including 3 CNNs) are
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
We would like to acknowledge the support that NVIDIA provided us through the GPU Grant Program. We used a donated TitanX GPU to train CNNs used in this work.
References (66)
- et al.
Hybrid human skin detection using neural network and K-means clustering technique
Applied Soft Computing
(2015) - et al.
Human skin detection through correlation rules between the YCb and YCr subspaces based on dynamic color clustering
Computer Vision and Image Understanding
(2017) - et al.
A skin tone detection algorithm for an adaptive approach to steganography
Signal Processing
(2009) - et al.
Statistical skin color detection method without color transformation for real-time surveillance systems
Engineering Applications of Artificial Intelligence
(2012) - et al.
A survey of skin-color modeling and detection methods
Pattern Recognition
(2007) - et al.
Spatial-based skin detection using discriminative skin-presence features
Pattern Recognition Letters
(2014) - et al.
Color based skin classification
Pattern Recognition Letters
(2012) - et al.
Recent advances in visual and infrared face recognition – A review
Computer Vision and Image Understanding
(2005) - et al.
Naked image detection based on adaptive and extensible skin color model
Pattern Recognition
(2007) - et al.
Skin detection by dual maximization of detectors agreement for video monitoring
Pattern Recognition Letters
(2013)
Objective evaluation of approaches of skin detection using ROC analysis
Computer Vision and Image Understanding
Pruning training sets for learning of object categories
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
Real-time tracking of multiple skin-colored objects with a possibly moving camera
Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SegNet: A deep convolutional encoder-decoder architecture for image segmentation
IEEE Transactions on Pattern Analysis and Machine Intelligence
SFA: A human skin image database based on FERET and AR facial images
Region-based and content adaptive skin detection in color images
International Journal of Pattern Recognition and Artificial Intelligence
Skin color modeling for face detection and segmentation: A review and a new approach
Multimedia Tools and Applications
A skin detector based on neural network
Encoder-decoder with atrous separable convolution for semantic image segmentation
Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Detector adaptation by maximising agreement between independent data sources
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
Partially-supervised learning from facial trajectories for face recognition in video surveillance
Information Fusion
Robust skin segmentation using color space switching
Pattern Recognition and Image Analysis
Automatic skin segmentation for gesture recognition combining region and support vector machine active learning
Face recognition in hyperspectral images
IEEE Transactions on Pattern Analysis and Machine Intelligence
Face detection in color images
IEEE Transactions on Pattern Analysis and Machine Intelligence
Human skin detection in images by MSER analysis
A dynamic skin detector based on face skin tone color
Adaptive skin color model to improve video face detection
Skin Detection Using Color, Texture and Space Information
Fourth International Conference on Fuzzy Systems and Knowledge Discovery
Statistical color models with application to skin detection
International Journal of Computer Vision
Fast propagation-based skin regions segmentation in color images
Self-adaptive algorithm for segmenting skin regions
EURASIP Journal on Advances in Signal Processing
Cited by (29)
Efficient hand segmentation for rehabilitation tasks using a convolution neural network with attention
2023, Expert Systems with ApplicationsDeep ensembles and data augmentation for semantic segmentation
2023, Diagnostic Biomedical Signal and Image Processing Applications with Deep Learning MethodsExploring the Potential of Ensembles of Deep Learning Networks for Image Segmentation
2023, Information (Switzerland)A color attention mechanism based on YES color space for skin segmentation
2023, Journal of Real-Time Image Processing