The Feature Extraction and Recognition of Phone Image Based on Robust Sparse Non-Negative Matrix Factorization

Enjun Yu, Qingwei Ye, Yue WuThe Open Automation and Control Systems Journal 2014 · DOI

Abstract

Sparse non-negative matrix factorization (SNMF) algorithms are effective for projecting image data and play a vital role in image matching and recognition. To enhance the effectiveness of SNMF in feature extraction from noisy image data, this paper introduces a noise term combined with the SNMF algorithm, resulting in a new sparse optimization objective function. This approach, termed Robust Sparse Non-Negative Matrix Factorization (RSNMF), guarantees the sparseness of extracted features while improving noise immunity. Applied to mobile phone interface recognition, the method utilizes RSNMF for feature extraction and a Support Vector Machine (SVM) for classification. Experimental results demonstrate that RSNMF enables large-scale data compression with high robustness, significantly improving recognition efficiency in automated testing environments.

Introduction

In the realm of automated mobile phone testing, systems must accurately identify the current interface (e.g., messaging, camera, or settings) to execute test scripts correctly. However, real-world image capture often introduces noise and data redundancy, which can degrade the performance of standard recognition algorithms. While Non-Negative Matrix Factorization (NMF) is a popular tool for dimensionality reduction, standard versions are notoriously sensitive to outliers and noise.

Existing Sparse NMF (SNMF) algorithms prioritize data compression but often ignore noise resistance. This paper addresses these challenges by proposing RSNMF, a robust variant designed to extract structural features from noisy high-dimensional images by explicitly modeling noise as a separate variable.

Robust Sparse Non-Negative Matrix Factorization (RSNMF)

The core innovation of RSNMF is the integration of a dedicated noise term ($E$) into the matrix factorization process. Instead of the standard decomposition $V \approx WH$, the proposed model is defined as:

$$V \approx WH + E$$

Where:

  • $V$ represents the original image data.
  • $W$ is the basis matrix (representing the "parts" of the image).
  • $H$ is the coefficient matrix.
  • $E$ is the noise term.

By using Kullback-Leibler (K-L) divergence and incorporating penalty terms for sparsity ($\lambda$) and noise ($\alpha$), the algorithm can "set aside" noise while focusing on the structural components of the interface. This allows the algorithm to maintain high recognition accuracy even when images are distorted by Gaussian white noise.

The Recognition Pipeline

The feature extraction and recognition process (FE-RSNMF) follows a structured pipeline. First, a set of training images is used to calculate a feature matrix. These features are integrated into a target matrix which is then used to train a multi-class SVM. For testing, the interface image is projected onto the learned basis to generate a feature vector, which the trained SVM then classifies.

Flowchart of the FE-RSNMF and SVM recognition process, illustrating the training and testing phases.

Experimental Evaluation

The proposed method was evaluated using a custom dataset of 150 mobile phone interface images across six categories: Main interface, Messaging, Text editing, Internet, Antivirus, and Camera. To simulate real-world conditions, Gaussian white noise with varying variances (0.002 to 0.006) was injected into the test set.

Example of a mobile phone interface used in the study. The highlighted regions represent sub-graphs or icons that serve as templates for feature extraction.

Fifteen "sub-graph" templates, such as home icons and settings gears, were extracted to serve as the basis for feature calculation. The performance of the RSNMF algorithm was compared against classical SNMF to measure its robustness against "abnormal point values" (noise).

Results and Discussion

The experimental results confirmed that RSNMF significantly outperforms standard SNMF in noisy environments. Key findings include:

  • Robustness: RSNMF effectively isolated noise into the $E$ term, preserving the structural integrity of the $WH$ reconstruction.
  • Efficiency: The algorithm achieved large-scale data compression by generating sparse intermediary matrices, reducing storage requirements without sacrificing classification accuracy.
  • Parameter Sensitivity: The "noise factor" $\alpha$ was found to be critical; if set too high, the algorithm reverts to standard SNMF behavior, losing its noise immunity.

Conclusion

RSNMF provides a powerful framework for feature extraction in noisy, part-based image data like mobile phone UIs. By treating noise as a variable rather than a pre-processing hurdle, the system achieves high recognition rates suitable for automated testing. Future work will focus on optimizing the iterative computation to handle extremely large-scale data in real-time and automating the selection of hyperparameters like the characteristic dimension $r$.

Summary updated

Abstract

Sparse non-negative matrix factorization algorithm can project image data effectively. It plays an important role in image matching and recognition. In order to improve the effectiveness of SNMF algorithm, which is used in feature extraction of image data with noises, we added a noise term and combined it with SNMF algorithm. Then, we proposed a new sparse optimization objective function and worked out its solution which can guarantee the sparseness of extracted feature and improve the algorithm's immunity against noise at the same time. We name this robust sparse nonnegative matrix factorization (RSNMF) algorithm. It is applied for feature extraction and recognition of phone image. The concept of interface image and sub-graph of mobile phone is created. The feature extraction of phone image is used in RSNMF. And the features are put in support vector machine to achieve classification recognition. Experimental results shows that not only phone image data can be large-scale compressed through RSNMF algorithm with good robustness, but also the recognition efficiency is improved by generating sparse matrix as an intermediary target matrix to classification.
BibTeX
@article{yu2014feature,
  author = {Enjun Yu and Qingwei Ye and Yue Wu},
  title = {{The Feature Extraction and Recognition of Phone Image Based on Robust Sparse Non-Negative Matrix Factorization}},
  journal = {The Open Automation and Control Systems Journal},
  year = {2014},
  doi = {10.2174/1874444301406011736},
}
Figure