Source-linked AI summary
A Fusion Approach for Efficient Human Skin Detection
Wei Ren Tan, Chee Seng Chan, Pratheepan Yogarajah, Joan Condell
TL;DR
Human skin detection must handle variation in skin colour and illumination while avoiding false detections and high computational cost. The paper combines a smoothed 2D histogram with a Gaussian model, uses eye-based person-specific refinement and dynamic thresholds, and reports effective, robust detection without training across varied conditions. Its main scope boundary is dependence on accurate eye detection.
Problem
Existing skin-detection methods can produce false skin detections, struggle with ethnic and illumination variation, and require substantial training or computation.
Method
The approach uses an eye detector, online dynamic thresholds, smoothed 2D-histogram and Gaussian distribution models, and product-rule fusion.
Results
Experiments report better accuracy and robustness than state-of-the-art methods across different backgrounds, illumination conditions, and ethnicities.
Takeaways & Limitations
The fusion framework provides a training-free approach with potential application to tasks such as gesture analysis.
Takeaways & Limitations
Performance depends greatly on the eye detector; false face regions can produce false dynamic thresholds and false skin classifications.
Abstract
from arXiv · showhide
A reliable human skin detection method that is adaptable to different human skin colours and illu- mination conditions is essential for better human skin segmentation. Even though different human skin colour detection solutions have been successfully applied, they are prone to false skin detection and are not able to cope with the variety of human skin colours across different ethnic. Moreover, existing methods require high computational cost. In this paper, we propose a novel human skin de- tection approach that combines a smoothed 2D histogram and Gaussian model, for automatic human skin detection in colour image(s). In our approach an eye detector is used to refine the skin model for a specific person. The proposed approach reduces computational costs as no training is required; and it improves the accuracy of skin detection despite wide variation in ethnicity and illumination. To the best of our knowledge, this is the first method to employ fusion strategy for this purpose. Qualitative and quantitative results on three standard public datasets and a comparison with state-of-the-art methods have shown the effectiveness and robustness of the proposed approach.
I. INTRODUCTION
Existing skin-detection methods face false detections across ethnicities, complex backgrounds, and illumination, while often requiring substantial training data. The paper proposes a training-free fusion framework combining a smoothed 2D histogram and Gaussian model, with dynamic thresholds and product-rule fusion.
- Applications: Skin detection supports applications including face tracking, gesture analysis, image retrieval, and human-computer interaction by reducing the search space for relevant objects.Skin-colour regions are also described as indicators of human presence in images and video.
- Related work: Existing approaches include thresholding, multilayer perceptrons, Bayesian classifiers, and random forests, but can trade precision against computational complexity.Some Bayesian approaches require very large training collections, including 2 billion pixels from 18,696 web images.
- Motivation: Skin colour varies across ethnicities, and fixed or pre-learned thresholds can fail under changing illumination and backgrounds.The cited methods are also described as suffering false skin detection and narrow robustness under luminance-invariant representations.
- Contribution: The proposed method uses online dynamic thresholds, eliminating a prior training stage for skin detection.The approach is presented as an advantage over state-of-the-art solutions.
- Contribution: It models skin and non-skin distributions with a smoothed 2D histogram and Gaussian model, respectively, then combines them using a product-rule fusion strategy.The paper identifies this as its first use of fusion for colour-image skin detection.
III. OUR METHOD
The framework detects faces, estimates dynamic skin thresholds, models skin and non-skin distributions with two features, and fuses them to produce skin-detection results.
- Face and threshold processing: The framework first obtains face regions with an eye-detector-related approach and calculates dynamic skin thresholds on those regions.The detected face provides the region used for threshold estimation.
- Distribution modelling: It introduces a smoothed 2D histogram and Gaussian model to represent skin and non-skin distributions, respectively.
- Fusion: A product-rule fusion framework combines the two features to obtain improved skin-detection results.
A. Pre-processing
The pre-processing stage detects human eyes, uses their geometry to define an elliptical face region, and removes non-smooth facial regions with Sobel-based processing.
- Human eyes are first located in each image to provide coordinates for subsequent face-region extraction.
- An elliptical mask defines the face region around the eyes, with minor and major axes of 1.6D and 1.8D, where D is the inter-eye distance.The ellipse centre is also the eyes’ symmetry point.
- Sobel edge detection removes non-smooth regions such as eyes, eyebrows, and mouth areas from the detected face regions.The detected edge pixels are then dilated to obtain the final non-smooth regions.
- The resulting image contains only the processed face regions used for later skin modelling.
B. Colour Space
The method uses log opponent chromaticity and an online smoothed 2D histogram to adapt skin thresholds to individual face images rather than relying on fixed boundaries.
- Log opponent chromaticity is selected because colour opponency is perceptually relevant and logarithms convert illumination changes into coordinate translations.
- Skin appearance varies with ethnicity, illumination, background, and camera characteristics, making fixed or pre-learned skin boundaries unsuitable.
- An online dynamic approach calculates skin threshold values from face images, assuming a person’s face and body share the same colours.
- The method replaces a 1D histogram with a smoothed 2D histogram using two channels from the selected colour representation.The 2D histogram feature vector Z is formed from I and By.
- The histogram classifier labels a pixel as skin when Z(I_n, By_n) > 20 and as non-skin when Z(I_n, By_n) ≤ 20.
2) Gaussian Model:
The Gaussian model represents the threshold skin distribution with elliptical Gaussian probability functions and classifies pixels using their geometry relative to the model.
- The Gaussian model describes complex-shaped skin-colour distributions using elliptical Gaussian joint probability functions.
- The colour vector H consists of (I, By), while λ contains mixing weights, mean vectors, and diagonal covariance matrices.
- Gaussian detection uses the model centre μ and calculates each pixel’s distance d and angle τ relative to that centre.The pixel coordinate is represented as a point in the colour plane, and these quantities are calculated from its position.
- The Gaussian boundary distances along the x- and y-axes are determined from the model centre, variances, and the pixel’s angle.
- The Gaussian skin-detection result Dgmm is obtained from the geometric quantities defined for the elliptical model.
D. Fusion Strategy
The fusion strategy combines the smoothed 2D histogram and Gaussian model so that both representations contribute to skin-versus-non-skin classification.
- The method fuses the smoothed 2D histogram and Gaussian model into a combined representation for skin detection.
- Each model produces matching results, and the two results vote on whether pixels are skin or non-skin.
- The selected fusion rule Γ is the product operator ⊗, under an assumption that the individual features are independent.
IV. EXPERIMENTS
The experiments evaluate the proposed approach under different fusion strategies and colour spaces, and compare it with state-of-the-art methods using qualitative and quantitative performance.
- IV. EXPERIMENTS: Experiments assess fusion strategy, colour-space choice, and comparisons with state-of-the-art methods.Quantitative analysis is limited to the Stottinger dataset because ground-truth videos are available only there.
A. Datasets
The evaluation uses three public datasets covering varied cameras, colour enhancements, illuminations, and amateur photographs. Qualitative detection results compare the proposed method with two prior methods across these datasets.
- A. Datasets: Experiments use Pratheepan, ETHZ PASCAL, and Stottinger public datasets.The Pratheepan dataset contains randomly downloaded images captured with different cameras, colour enhancements, and illuminations; ETHZ PASCAL contains 549 PASCAL VOC 2009 images.
- A. Datasets: The proposed method qualitatively detects skin more accurately and robustly than and under complex backgrounds, high illumination, varied cameras, and different ethnicities.In the cited examples, competing methods miss skin regions, whereas the proposed approach captures almost all skin regions with the least noise.
C. Comparison between different colour spaces
The study compares seven colour-feature combinations and selects IBy because it provides higher true positive and lower false negative rates than the alternatives, while fusion outperforms single-feature models.
- C. Comparison between different colour spaces: Seven feature-vector combinations are evaluated: IBy, HS, HV, SV, Y Cb, Y Cr, and CbCr.The comparisons use images from the Pratheepan and ETHZ datasets, with quantitative analysis on Stottinger.
- C. Comparison between different colour spaces: IBy is selected because it has a higher true positive rate and lower false negative rate than HS, although their results are comparable.IBy shows better true positive rate than the other evaluated combinations.
- D. Fusion strategy results: Fusion of the smoothed 2D histogram and Gaussian model achieves the highest Accuracy and F-score among the compared approaches.The fusion strategy also has a lower false positive rate than either single-feature approach.
E. Quantitative Analysis
Quantitative and cross-dataset comparisons indicate that the proposed method maintains acceptable performance without training while handling illumination and background variation, unlike a trained random forest that requires substantial training time.
- E. Quantitative Analysis: The Stottinger evaluation extracts 2,985 frames from seven videos, while Random Forest trains on 1,990 frames and approximately 3 million randomly selected pixels.Fifteen trees are trained, with each tree using 70% of the pixels.
- E. Quantitative Analysis: On Pratheepan images, Random Forest trained on Stottinger does not work well, whereas the proposed method maintains skin-segmentation quality.Improving Random Forest accuracy would require larger training samples or more trees, increasing computational power and training time.
- E. Quantitative Analysis: The proposed method achieves an acceptable score compared with Random Forest on Stottinger despite requiring no training, while coping with illumination and complex-background variation.Table III reports quantitative accuracy and F-score comparisons using the Stottinger dataset.
F. Discussions
The method performs well across different backgrounds, illumination conditions, and ethnicities, but its success depends strongly on the eye detector used during preprocessing.
- A false face region produces false dynamic thresholds and can classify non-skin regions as skin.
- The proposed approach outperforms state-of-the-art methods in accuracy across different background models, illumination conditions, and ethnicities.
- The authors identify reliance on eye detector algorithms as a drawback and propose improving preprocessing in future work.