An Online Tea Fixation State Monitoring Algorithm Based on Image Energy Attention Mechanism and Supervised Clustering (IEAMSC)

Zhiyong Yu, Jin Wang, Tao Zheng, Guodong LuSensors 2020 · DOI

TL;DR

This paper introduces IEAMSC, a novel algorithm for real-time monitoring of tea fixation quality using image analysis. By combining an adaptive filtering approach with an image energy attention mechanism and supervised clustering, IEAMSC effectively extracts key features and achieves high accuracy in determining the tea fixation state, outperforming existing methods.

What Problem Does This Paper Address?

The quality of tea, particularly its flavor and medicinal properties, is heavily influenced by the "fixation" process, which halts oxidation. Current methods for monitoring this crucial step often rely on manual judgment by experienced workers or fixed temperature/time settings, which are insufficient for automated production lines. Existing automated methods, like spectral analysis or electronic noses, are either expensive, time-consuming, or lack direct correlation with the fixation state. Image-based methods have also faced challenges with background noise, illumination variations, and computational complexity, making real-time, accurate monitoring difficult. This paper aims to bridge this gap by developing an image-based algorithm that is robust, efficient, and suitable for online monitoring.

Methods

The proposed IEAMSC algorithm comprises several key steps:

  • Adaptive Filtering: This initial stage removes ineffective background and overlapping areas in the tea images. Unlike fixed threshold methods, it dynamically adjusts thresholds based on the image's energy distribution, preserving essential tea features.
  • Energy Extraction: The algorithm quantizes image energy by counting pixels for each gray level across the Red, Green, and Blue (RGB) color channels. This captures the overall brightness and color characteristics related to fixation.
  • Image Energy Attention Mechanism: This step focuses on the most prominent color features within the filtered image. It reduces data dimensionality by identifying the grayscale range with concentrated energy and calculating average attention values for each color channel.
  • Data Mapping: The extracted energy features are mapped to a 2D plane, creating a uniform representation that highlights differences between images and further reduces dimensionality.
  • Supervised Clustering Training: Using labeled fixation data, the algorithm trains a clustering model. It identifies "feature clustering centers" for different fixation states by averaging the mapped data points. The core idea is that as fixation progresses, the clustering centers for the three color channels converge.
  • Decision Function and Online Monitoring: A decision function is constructed based on the Euclidean distance between these clustering centers. By comparing the distances of new, unseen tea images to these established standards, the algorithm can accurately determine if the fixation state is normal or abnormal in real-time. A novel observation is that the triangle formed by the three color channel clustering centers shrinks as fixation time increases.

Datasets

The study utilized fresh tea leaves from two categories: Wu Niu Zao and Ying Shuang.

  • Training Data: 2800 images were collected for the Wu Niu Zao tea, with 700 images for each fixation time point (2, 4, 6, and 8 minutes).
  • Test Data: 1200 images were initially collected for the Ying Shuang tea. A data enhancement technique was used to expand this to 3000 images for testing.

The paper does not explicitly mention the use of publicly available benchmark datasets for this specific task.

Key Insights and Ideas

  • The "Shrinking Triangle" Phenomenon: A significant discovery is that as the tea fixation time increases, the Euclidean distance between the energy clustering centers of the R, G, and B channels progressively decreases. This visually manifests as a shrinking triangle formed by these centers, indicating a convergence of features and a more stable fixation state. This geometric trend provides a novel way to assess fixation quality.
  • Adaptive Filtering for Robustness: The use of adaptive filtering, rather than fixed thresholds, is crucial for handling variations in background, illumination, and overlapping tea leaves. This allows the algorithm to focus on the intrinsic features of the tea itself.
  • Energy Attention for Feature Salience: The attention mechanism effectively highlights the most informative color features related to fixation, reducing noise and computational load while preserving critical information.
  • Supervised Clustering for State Definition: By training on labeled data, the algorithm learns to define distinct "states" of fixation, allowing for precise classification of online monitoring results.

Key Results

The IEAMSC algorithm demonstrated strong performance:

  • High Accuracy: The proposed algorithm achieved an accuracy of 95% and a precision of 94%, significantly outperforming other image-based methods like GLCM+SVM (74% accuracy), HOG+SVM (61% accuracy), HSV+SVM (66% accuracy), and LASC (64% accuracy).
  • Robustness: The algorithm showed improved resilience to background noise and illumination variations compared to other methods.
  • Efficiency: The method is designed for online monitoring, implying a reasonable processing speed suitable for automated production lines.

Limitations and Future Work

  • Ambient Light Sensitivity: While the algorithm is robust, the experimental results indicate that ambient light can still influence the data. The authors suggest using multi-angle annular light sources and shielding ambient light for even better results.
  • Uniform Heating: Uneven heating during the fixation process can lead to color differences within the same fixation time, potentially affecting test results.
  • Future Work: The authors suggest integrating all feature information (from Table 1) for more comprehensive decision functions. They also propose developing the system to not only signal the fixation state but also provide accurate fixation time recommendations. Further research could explore more sophisticated decision functions and potentially incorporate temporal analysis of image sequences.

Summary updated

Abstract

This study aimed at the shortcomings of existing fixation algorithms that are image-based only, and an effective tea fixation state monitoring algorithm was proposed. An adaptive filtering algorithm was used to automatically filter the ineffective information. Using the energy extractor, the complete energy information of each fixation image was extracted. The image energy attention mechanism was used to identify the prominent features, and based on these, the energy data was mapped to generate the data points as the training data. The cluster idea was adopted, and the training data feed the features trainer. The trend center data of the tea processing energy clustering was generated from different color channels. The corresponding decision function was designed which is based on the distance of the cluster center. The fixation degree of each monitoring image set was measured by the decision function. The Euclidean distance of the energy clustering center of the three channels with the same fixation time progressively approached. The triangle formed by these three points had a trend of gradually shrinking, which was first discovered by us. The detection results showed high accuracy compared with the common classification algorithms. It indicates that the algorithm proposed has positive guiding and reference significance.
BibTeX
@inproceedings{yu2020online,
  author = {Zhiyong Yu and Jin Wang and Tao Zheng and Guodong Lu},
  title = {{An Online Tea Fixation State Monitoring Algorithm Based on Image Energy Attention Mechanism and Supervised Clustering (IEAMSC)}},
  booktitle = {Sensors},
  year = {2020},
  doi = {10.3390/s20154312},
}
Figure