Feature extraction of 3D Chinese rose model based on color and shape features

Jin’fei Liu, Shu’li Mei, Tao Song, Hong’hao LiuFrontiers in Plant Science 2022 · DOI

This research paper, published in Frontiers in Plant Science, presents a sophisticated approach to botanical classification by moving from flat 2D images to complex 3D models. By fusing geometric shape descriptors with color data and a novel hierarchical sampling method, the authors have developed a system capable of identifying Chinese rose varieties with high precision.

TL;DR

Researchers have developed a new 3D feature extraction method that combines color and shape data to classify Chinese rose varieties with 87% accuracy. By using hierarchical sampling and Random Forest-based feature ranking, the study overcomes the perspective and lighting limitations of traditional 2D flower recognition.

What Problem Does This Paper Address?

Flower classification is essential for fields ranging from pharmacology to automated agriculture. However, traditional computer vision relies heavily on 2D images, which are "vulnerable" to environmental factors like lighting, background clutter, and the "one visual angle" problem—where a flower looks different depending on how the camera is positioned.

While 3D models provide "omnidirectional" information, they are computationally expensive and difficult to describe mathematically. Standard 3D shape descriptors often fail to capture the delicate, small-area mesh details of organic structures like rose petals. The authors sought to bridge this gap by creating a descriptor specifically tuned to the unique morphological attributes of the Chinese rose.

Methods

The authors proposed a multi-pronged feature extraction pipeline:

  1. Shape Distribution: They utilized six standard geometric functions (D1, D2, D3, A3, D4, C1) to measure distances, angles, and volumes between random points on the model's surface.
  2. New Morphological Descriptors: They introduced two specific features:
  3. Sharpness (S7): The ratio of maximum to minimum distance from the center to the boundary, representing flower "fragility."
    • Contour (P8): The ratio of the flower’s volume to its bounding sphere, representing "fullness" or "openness."
  4. Hierarchical Sampling: To capture the complex internal structure of the rose, they divided the 3D model into three layers—In (stamen), Mid (internal petals), and Out (outer contour)—ensuring that features were extracted from the entire volume, not just the surface.
  5. Color Integration: Color features were extracted from the RGB space, including a filtering step to remove interference from green leaves.
  6. RF-OOB Weighting: They used the Random Forest Out-of-Bag (RF-OOB) method to rank the importance of these nine features. This allowed them to apply an "exponential distribution weight" to the most critical features during classification.

Datasets

The study utilized a custom dataset provided by the College of Horticulture at China Agricultural University:

  • Subjects: 15 varieties of Chinese roses (e.g., 'Golden Phoenix', 'Lady Figaro', 'Scarlet Queen').
  • Sample Size: 40 3D models per variety, totaling 600 models.
  • Colors: Eight distinct color categories (Red, Yellow, Pink, Black/Red, Blue, Green, White, Orange).

Key Insights and Ideas

  • Color is Still King: Despite the move to 3D geometry, the RF-OOB analysis revealed that color (C9) remains the most important feature for distinguishing rose varieties.
  • The "Fullness" Factor: The new Contour (P8) feature ranked second in importance, suggesting that the overall "plumpness" of a bloom is a primary differentiator in 3D space.
  • Layering is Essential: By stratifying the 3D model into three layers, the researchers could capture the stamen and internal petal morphology that is often lost in simplified surface-mesh analysis.
  • Smart Weighting Beats Averaging: The study found that giving more mathematical "weight" to the most important features (exponential weighting) significantly outperformed a simple average of all features.

Key Results

  • Accuracy: The average classification accuracy reached 87%.
  • Weighting Performance: The exponential distribution weight method achieved a 90% average accuracy in some tests, which was 3% higher than the average distribution weight method.
  • Retrieval Effectiveness: In similarity retrieval tests (finding similar roses in a database), the fused model (RF-OOB-2) significantly outperformed models that ignored color or used standard 3D descriptors alone. Subjectively, ignoring color led to "poor" results where orange and pink roses were returned for a white rose query.

Limitations and Future Work

While successful, the authors note several areas for growth:

  • Species Diversity: The current model is specifically tuned to Chinese roses. Its performance on other flower species with radically different geometries (like orchids or lilies) is unknown.
  • Dataset Scale: A total of 600 models is relatively small for deep learning contexts; expanding the library would improve robustness.
  • Technological Integration: The authors suggest that future work should explore the application of AI and Virtual Reality (VR) to better visualize and evaluate these 3D models in real-time botanical research.

Summary updated

Abstract

Flower classification is of great importance to the research fields of plants, food, and medicine. Due to more abundant information on three-dimensional (3D) flower models than two-dimensional 2D images, it makes the 3D models more suitable for flower classification tasks. In this study, a feature extraction and classification method were proposed based on the 3D models of Chinese roses. Firstly, the shape distribution method was used to extract the sharpness and contour features of 3D flower models, and the color features were obtained from the Red-Green-Blue (RGB) color space. Then, the RF-OOB method was employed to rank the extracted flower features. A shape descriptor based on the unique attributes of Chinese roses was constructed, χ<sup>2</sup> distance was adopted to measure the similarity between different Chinese roses. Experimental results show that the proposed method was effective for the retrieval and classification tasks of Chinese roses, and the average classification accuracy was approximately 87%, which can meet the basic retrieval requirements of 3D flower models. The proposed method promotes the classification of Chinese roses from 2D space to 3D space, which broadens the research method of flower classification.
BibTeX
@inproceedings{liu2022feature,
  author = {Jin’fei Liu and Shu’li Mei and Tao Song and Hong’hao Liu},
  title = {{Feature extraction of 3D Chinese rose model based on color and shape features}},
  booktitle = {Frontiers in Plant Science},
  year = {2022},
  doi = {10.3389/fpls.2022.1042016},
}
Figure