Complex Background Palm Segmentation using SVM

Peng Gao, Liang LiInternational Journal of Software Engineering and Its Applications 2015 · DOI

TL;DR

This paper introduces a novel method for segmenting hands from complex backgrounds in images, crucial for robust contactless palmprint recognition. By combining texture and "haze" features extracted using Laws masks in the YCbCr color space, and incorporating neighborhood information, their SVM-based approach achieves high accuracy (94.96%) even under challenging illumination and varied backgrounds.

What Problem Does This Paper Address?

Traditional palmprint recognition systems often rely on specialized scanners that require direct hand contact. While contactless methods offer advantages in hygiene and flexibility, they face a significant challenge: accurately segmenting the hand from a complex, variable background. Existing methods struggle with changing illumination and diverse environments, leading to reduced recognition accuracy. This paper aims to tackle this by developing a robust hand segmentation technique that can effectively isolate the hand region even when the background is cluttered and lighting conditions vary, paving the way for more reliable contactless biometric systems.

Methods

The core of the proposed method lies in a multi-stage feature extraction and classification process:

  • Color Space Transformation: Images are converted to the YCbCr color space. This is a key choice as YCbCr is known for its robustness to illumination variations compared to RGB.
  • Texture and Haze Feature Extraction: Laws masks, a set of 3x3 convolution kernels designed to capture different texture properties (level, edge, spot), are applied to the YCbCr channels. Specifically, all nine Laws masks are used on the intensity (Y) channel, while only the first Laws mask (L3) is applied to the chrominance (Cb and Cr) channels. This process generates 11 texture features per patch.
  • Neighborhood Information: Recognizing that a patch's classification is influenced by its surroundings, the method incorporates features from neighboring patches. Specifically, it uses the color features of a central patch and its four immediate neighbors.
  • Support Vector Machine (SVM) Classification: An SVM classifier is employed for the binary classification task: distinguishing between hand regions (+1) and non-hand regions (-1). The SVM is trained on the extracted features and neighborhood information. A Radial Basis Function (RBF) kernel is used for handling non-linearly separable data.

Datasets

  • Custom Dataset: The authors collected 120 images from 7 individuals. These images were captured using a MOTO ME865 phone with an 8MP camera, featuring various illumination conditions and random backgrounds.
  • Training/Testing Split: 70% of the collected images were used for training, and the remaining 30% for testing.

Key Insights and Ideas

  • YCbCr for Robustness: The choice of YCbCr color space is crucial for handling illumination variations, a common problem in real-world scenarios.
  • Laws Masks for Texture and Haze: Laws masks effectively capture both fine-grained texture details and broader "haze" or low-frequency information, providing rich feature representations.
  • Neighborhood Context is Key: Incorporating information from neighboring patches significantly improves segmentation accuracy, as it leverages spatial context and reduces ambiguity.
  • SVM's Power: SVM proves to be a powerful classifier for this task, capable of handling high-dimensional feature spaces and achieving good generalization.

Key Results

  • High Accuracy: The proposed method achieved an average accuracy of 94.96% on their custom dataset.
  • Robustness: The algorithm demonstrated robustness to changing illumination and complex environments.
  • Qualitative Improvement: Figure 8 shows that the proposed method produces smoother and clearer edge lines for the palm compared to other methods, indicating better qualitative segmentation.
  • Feature Importance: Feature analysis revealed that the 1st, 10th, and 11th features (likely related to level, and specific combinations of edge/spot filters) were most correlated with the target.
  • Neighborhood Benefit: Experiments showed that using neighborhood information improved accuracy, with an average accuracy of 91.87% for features with neighbors versus 90.37% without.

Limitations and Future Work

  • Dataset Size: While 120 images were collected, the authors acknowledge the need for larger and more diverse datasets for more comprehensive validation.
  • Generalization: The paper focuses on palmprint segmentation. Future work could explore extending this method to other biometric modalities or general object segmentation in complex scenes.
  • Computational Cost: While not explicitly detailed as a limitation, the multi-stage feature extraction and SVM classification might have computational implications for real-time applications on resource-constrained devices. Further optimization could be explored.

Summary updated

Abstract

An efficient method is proposed to enhance complex background segmentation for preprocessing in palmprint recognition. In this paper, we integrate texture and haze features by applying Laws masks to the image represented in YCbCr color space, and use these features of a patch with its neighborhood information to determine hand and non-hand regions by support vector machine (SVM). Compared with other methods, our algorithm demonstrated robustness to changing illumination and a complex environment and we obtain a relatively clear hand shape with an average accuracy of 94.96%. The images in our experiments are taken with popular mobile phones in our laboratory.
BibTeX
@article{gao2015complex,
  author = {Peng Gao and Liang Li},
  title = {{Complex Background Palm Segmentation using SVM}},
  journal = {International Journal of Software Engineering and Its Applications},
  year = {2015},
  doi = {10.14257/ijseia.2015.9.5.33},
}
Figure