Automatic recognition of wild flowers

Takeshi Saitoh, Toyohisa KanekoSyst. Comput. Jpn 2003 · DOI

TL;DR

This paper introduces an automatic system for recognizing wild flower species using digital photographs of flowers and leaves. By extracting features from both visual elements and color, the system achieves a high recognition rate, demonstrating the potential of computer vision for botanical identification.

What Problem Does This Paper Address?

Identifying wild flower species can be a challenging task for amateurs. Traditional methods rely on botanical classifications and detailed guides, which are often time-consuming and require specialized knowledge. The proliferation of digital cameras and portable computers presents an opportunity to develop automated identification systems. While research exists for recognizing objects like lung cancer or faces, the recognition of plants, particularly wild flowers, has been less explored. This paper aims to bridge this gap by creating a system that handles the entire process from image acquisition to recognition of living wild flowers.

Methods

The proposed system employs a multi-stage approach:

  • Background Separation: To isolate the flower and leaves from the background, the authors utilize a clustering method in RGB color space. This involves:
  • Initial Cluster Determination: Identifying two initial clusters (background and objective) by analyzing the distribution of pixels in RGB space and selecting a "seed" based on maximal frequency in small regions, while excluding neighboring regions to avoid bias.
  • K-Means Clustering: Applying the k-means algorithm to group pixels into clusters based on their RGB values.
  • Cluster Integration: A three-step process to refine the clusters:
  • Primary Integration: Merging small clusters based on color similarity to reduce noise and minor variations.
  • Second Integration: Differentiating between background and objective clusters by analyzing "edge force" (differences in RGB values at cluster boundaries), assuming the objective is centered in the image.
  • Third Integration: Removing shadows by comparing the average intensity of objective clusters with the background intensity.
  • Feature Extraction: Once the objective (flower and leaves) is extracted, a set of features are computed:
  • Flower Features (10 total): * Shape (4 features): Derived from a one-dimensional waveform representing the distance from the flower's center to its contour. These include petal shape (width/height ratio), number of petals, moment (representing complexity), and roundness.
  • Color (6 features): Based on a two-dimensional histogram of the flower's HS color space, capturing the distribution and frequency of the two dominant colors.
  • Leaf Features (11 total): * Shape (8 features): Including serration rate, aspect ratio, moment, roundness, bias in the center of gravity, stem angle, tip angle, and leaf structure (pinnate/ternate).
  • Color (3 features): Similar to flower color features, focusing on the dominant green color distribution in HS space.
  • Recognition: A piecewise linear discriminant function is used for classification, with normalized features as input. The "leave-one-out" method is employed for evaluation to estimate the true recognition rate from a limited dataset.

Datasets

The experiments were conducted on 20 sets of images for each of the 34 species of wild flowers growing around the authors' university campus. These flowers bloom in spring and early summer. The paper includes an appendix with images of these 34 species.

Key Insights and Ideas

  • Flower Features Dominate: The study found that flower features are significantly more useful for recognition than leaf features, contributing more to the overall accuracy.
  • Color is Crucial for Flowers: Among the effective flower features, color information played a vital role, with three out of the six most useful features being color-based.
  • Overall Shape Matters: For flower shape features, those representing the overall shape (moment and roundness) were more impactful than detailed petal structures.
  • Clustering for Robust Extraction: The multi-stage clustering and integration process proved effective in accurately separating the objective (flower/leaf) from complex backgrounds, including shadows.
  • Limited Features Suffice: A high recognition rate was achieved with a relatively small subset of features (seven or eight), suggesting that the defined features are well-suited for distinguishing species.

Key Results

  • Overall Recognition Rate: Using all 21 features, the system achieved a recognition rate of 96.0%.
  • Most Effective Features: The authors identified six flower features and two leaf features as the most useful. Using these eight features, the recognition rate improved to 96.8%.
  • Background Extraction Success: The background extraction rate was high, 98.5% for flower images and 99.1% for leaf images.
  • Feature Importance: The paper demonstrates that even with size information omitted, a high recognition rate was achieved, highlighting the effectiveness of the chosen color and shape features.

Limitations and Future Work

  • Controlled Environment: The current system relies on images taken against a black cloth or paper. A significant future direction is to develop a method for extracting objectives from natural images without such controlled backgrounds.
  • Three-Dimensional Structures: The current features are primarily two-dimensional. Future work should explore incorporating three-dimensional structural information for more comprehensive botanical classification.
  • Seasonal and Locational Information: The experiment did not consider seasonal or collection site information. Evaluating the impact of adding these features is a promising avenue.
  • Genus-Level Misrecognition: The system struggled with distinguishing species within the same genus, and with flowers that have similar shapes and colors.
  • User Interface: The goal was to identify a single correct name. Future systems could display a list of similar results, allowing human eyes to make the final determination.
  • Dataset Size: The experiment used 20 sets of images for 34 species. Analyzing the recognition rate variations with a larger number of wild flower species is necessary.

Summary updated

BibTeX
@article{saitoh2003automatic,
  author = {Takeshi Saitoh and Toyohisa Kaneko},
  title = {{Automatic recognition of wild flowers}},
  journal = {Syst. Comput. Jpn.},
  year = {2003},
  doi = {10.1002/scj.10099},
}

Loading related content...

Figure