Classification of Tea Leaves Based on Fluorescence Imaging and Convolutional Neural Networks

Kaihua Wei, Bojian Chen, Zejian Li, Dongmei Chen, Guangyu Liu, Hongze Lin, Baihua Zhang

TL;DR

This paper introduces a novel tea classification method using fluorescence imaging induced by UV LEDs and deep learning (CNNs). By leveraging the unique fluorescence signatures of tea leaves, this approach achieves a remarkable 97.5% classification accuracy, outperforming traditional white-light imaging and offering a promising avenue for robust, smartphone-integrated tea quality assessment.

What Problem Does This Paper Address?

The global tea market faces challenges with the increasing prevalence of counterfeit or lower-quality products, making it difficult for consumers to distinguish genuine, high-grade teas. Traditional manual evaluation by experts is time-consuming, expensive, and subjective. While computer vision offers a more objective approach, existing methods often struggle with variations in illumination conditions, impacting prediction robustness. This paper aims to address this by exploring fluorescence imaging as a more stable and informative technique for tea classification, specifically investigating its efficacy when combined with deep learning models.

Methods

The core of this research lies in the combination of fluorescence imaging and deep learning.

  • Fluorescence Imaging: The authors utilized UV LEDs with a central wavelength of 370 nm to excite fluorescence in tea samples. This specific wavelength was chosen to capture fluorescence from both chlorophyll (red and far-red bands) and tea polyphenols (yellow band). Images were captured using a CMOS camera. For comparison, images were also taken under conventional white-light illumination.
  • Deep Learning Models: Two established Convolutional Neural Network (CNN) architectures were employed:
  • VGG16: A deep CNN known for its simplicity and effectiveness, characterized by repeated 3x3 convolution kernels and 2x2 max-pooling layers.
  • ResNet-34: A deep residual network designed to overcome the degradation problem in very deep networks by incorporating identity shortcut connections.
  • Transfer Learning: Both VGG16 and ResNet-34 models were pre-trained on the ImageNet dataset and then fine-tuned for the tea classification task.
  • Image Preprocessing: Key steps included:
  • Region of Interest (ROI) Extraction: To isolate the tea leaves and remove background noise.
  • Data Augmentation: To expand the limited dataset and prevent overfitting, employing techniques like brightening, darkening, flipping, rotation, padding, and noise injection.
  • Channel Analysis: Experiments were conducted using individual RGB channels, grayscale images, and the full RGB image to understand the contribution of different color information.
  • Evaluation Metrics: Standard metrics like accuracy, precision, and recall were used, with a macro-average method for multi-class classification.

Datasets

The study utilized a custom dataset comprising images of five different tea species: Anjibai (AJB), Maofeng (MF), West Lake Longjing (LJ), Huangjingui (HJG), and Tieguanyin (TGY). For each tea type, 76 fluorescence images and 76 white-light images were collected. The initial dataset consisted of 300 samples for training and 80 for testing. Through data augmentation, the training set was expanded to 3300 images. No public benchmark datasets for this specific fluorescence-based tea classification were mentioned.

Key Insights and Ideas

  • Fluorescence as a Discriminator: The study strongly suggests that fluorescence imaging provides richer, more discriminative features for tea classification compared to standard white-light imaging. This is likely due to the unique biochemical composition of different tea types, which manifest as distinct fluorescence patterns.
  • Illumination Robustness: Fluorescence imaging, particularly when excited by controlled UV LEDs, offers greater robustness against variations in ambient lighting conditions, a significant limitation of traditional computer vision methods.
  • RGB Channels are Crucial: While individual color channels (R, G, B) offer some discriminatory power, the full RGB image consistently yields superior classification accuracy. This indicates that all color channels contribute to capturing the complex spectral and shape features necessary for accurate tea identification.
  • VGG16's Edge: In this specific application, the VGG16 architecture demonstrated a slight but consistent performance advantage over ResNet-34 in terms of accuracy, precision, and recall. This highlights that architectural choices can still play a significant role, even with deep learning.

Key Results

  • Highest Accuracy: The fluorescence imaging approach achieved a peak classification accuracy of 97.5% using the VGG16 model with RGB images.
  • White-Light Performance: For comparison, white-light images yielded a maximum accuracy of 92.5% with VGG16.
  • Single-Channel Limitations: Classification accuracy using single RGB channels or grayscale images was significantly lower, peaking around 80% for fluorescence and 77.5% for white-light images.
  • Model Comparison: VGG16 consistently outperformed ResNet-34 by approximately 2% in accuracy, precision, and recall across both datasets.
  • Dataset Advantage: Fluorescence images provided a performance improvement of around 5% compared to white-light images for the same deep learning model.

Limitations and Future Work

  • UV LED Requirement: A current limitation is the need for an external UV LED device, which would need to be integrated into smartphone applications. The authors suggest exploring ways to minimize or remove this external device, potentially through LED modulation techniques to mitigate ambient light interference.
  • Smartphone Integration: While promising for smartphones, the current setup requires additional hardware. Future work aims to adapt the technology for smartphone use, possibly by leveraging existing smartphone light sources (though these may not be optimal for all fluorescence excitation).
  • Adulteration and Species Identification: The authors propose extending this work to identify tea leaf adulteration and to classify individual leaf species with higher granularity, potentially by using multiple excitation LEDs.
  • Channel Optimization: Further research could focus on optimizing channel selection and potentially employing more advanced filters to enhance prediction accuracy.

Summary updated

Abstract

The development of the smartphone and computer vision technique provides customers with a convenient approach to identify tea species, as well as qualities. However, the prediction model may not behave robustly due to changes in illumination conditions. Fluorescence imaging can induce the fluorescence signal from typical components, and thus may improve the prediction accuracy. In this paper, a tea classification method based on fluorescence imaging and convolutional neural networks (CNN) is proposed. Ultra-violet (UV) LEDs with a central wavelength of 370 nm were utilized to induce the fluorescence of tea samples so that the fluorescence images could be captured. Five kinds of tea were included and pre-processed. Two CNN-based classification models, e.g., the VGG16 and ResNet-34, were utilized for model training. Images captured under the conventional fluorescent lamp were also tested for comparison. The results show that the accuracy of the classification model based on fluorescence images is better than those based on the white-light illumination images, and the performance of the VGG16 model is better than the ResNet-34 model in our case. The classification accuracy of fluorescence images reached 97.5%, which proves that the LED-induced fluorescence imaging technique is promising to use in our daily life.
BibTeX
@inproceedings{wei2022classification,
  author = {Kaihua Wei and Bojian Chen and Zejian Li and Dongmei Chen and Guangyu Liu and Hongze Lin and Baihua Zhang},
  title = {{Classification of Tea Leaves Based on Fluorescence Imaging and Convolutional Neural Networks}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=FlOdqcnEwv},
}
Figure