Source-linked AI summary
Leaf Classification Using Shape, Color, and Texture Features
Abdul Kadir, Lukito Edi Nugroho, Adhi Susanto, Paulus Insap Santosa
TL;DR
The paper addresses leaf classification methods that often did not incorporate color information. It combines shape, vein, color, and texture features with a Probabilistic Neural Network and reports 93.75% accuracy on the 32-class Flavia dataset.
Problem
Prior plant-identification methods often omitted color information, while leaf classification can involve shape, color, vein, and texture features.
Method
The method extracts shape and vein features, color moments, lacunarity-based texture features, and normalized descriptors for a PNN classification system.
Results
93.75% accuracy was achieved by combining shape, color without kurtosis, vein, and texture features on the Flavia dataset.
Takeaways & Limitations
The combined-feature method gives slightly better accuracy than the original work’s 90,312% result.
Abstract
from arXiv · showhide
Several methods to identify plants have been proposed by several researchers. Commonly, the methods did not capture color information, because color was not recognized as an important aspect to the identification. In this research, shape and vein, color, and texture features were incorporated to classify a leaf. In this case, a neural network called Probabilistic Neural network (PNN) was used as a classifier. The experimental result shows that the method for classification gives average accuracy of 93.75% when it was tested on Flavia dataset, that contains 32 kinds of plant leaves. It means that the method gives better performance compared to the original work.
I. INTRODUCTION
The paper addresses plant-leaf identification by combining shape, vein, color, and texture information with a PNN classifier, motivated by prior methods that often omitted color. Tested on the 32-class Flavia dataset, the method improves performance over Wu’s result.
- Prior plant-identification methods used features such as aspect ratio, leaf dent, perimeter-to-diameter ratio, invariant moments, and veins.
- Earlier work included 32 plant kinds in the Flavia dataset, which was released for academic research and later used for method comparisons.
- Color, texture, and shape are common application features, but some leaf-identification studies used only part of them or omitted color information.
- Lacunarity is a fractal measure of data non-homogeneity or lumpiness, and the paper uses a modified definition for texture features.
- The proposed system extracts PFT Fourier descriptors, geometric, color-moment, vein, and lacunarity-based texture features, then classifies leaves with a PNN.
- The study evaluates the proposed method on Flavia and reports improved identification performance compared with Wu’s result.
II. FEATURE EXTRACTION
The feature-extraction stage obtains leaf descriptors before classification; the extracted features cover shape, color, veins, and texture.
- Features are extracted from the leaf’s shape, color, veins, and texture for use in the identification system.
A. Shape Features
Shape representation combines geometric descriptors with Polar Fourier Transform descriptors to characterize leaf geometry, irregularity, and invariant structure.
- Geometric shape features include slimness and roundness, while Fourier descriptors are derived from the Polar Fourier Transform.
- Slimness uses leaf width and length as its geometric parameters.
- Roundness, also called compactness, is defined from the leaf image’s area and contour perimeter.
- Dispersion represents irregular shape by comparing enclosing and contained circle radii, increasing as the region spreads.It is intended for irregular leaves but is insensitive to slight discontinuities such as cracks.
- PFT descriptors are invariant to translation, scaling, and rotation, supporting shape comparison despite positional, size, or orientation changes.
- PFT computation converts the image to polar space around the shape centroid, retains coefficient magnitudes for rotation invariance, and normalizes magnitudes for scale invariance.
B. Color Features
Color moments represent color features by summarizing the RGB image planes with statistical descriptors.
- Color moments are used to characterize color images through statistical features such as mean, standard deviation, skewness, and kurtosis.
- For RGB images, the color descriptors are extracted separately from the R, G, and B planes.
C. Vein Features
Vein features are extracted through morphological opening on grayscale leaf images, then summarized using three ratios based on vein and leaf pixel counts.
- C. Vein Features: Morphological opening extracts vein features from a grayscale image using flat, disk-shaped structuring elements with example radii 1, 2, and 3.The processed image is subtracted from the margin to isolate vein information.
- C. Vein Features: Three vein features are calculated from the extracted vein regions.The supplied equation passage presents the three feature definitions, while the accompanying text identifies their variables.
- C. Vein Features: V1, V2, and V3 denote vein features, while A1, A2, and A3 count vein pixels and A counts total pixels in the leaf region.These variables connect the feature formulas to pixel counts in the vein and leaf areas.
D. Texture Features
The paper uses lacunarity as a texture measure because fractal dimension is considered inadequate for image texture, and applies the formulas across color and grayscale channels.
- D. Texture Features: Lacunarity is introduced as a fractal measure that can distinguish structures sharing the same fractal dimension.The paper contrasts this with the view that fractal dimension is not a good texture descriptor for images.
- D. Texture Features: 12 texture features are extracted by applying Eq. 15 with p values 2, 4, and 6 to RGB color values and grayscale intensity.The implementation uses the specified p settings for each of the red, green, blue, and intensity representations.
E. Feature Normalization
Feature normalization is applied before classifier design to handle differing feature ranges and place feature values within predetermined ranges.
- E. Feature Normalization: Normalization is used before designing the classifier when feature values vary across different dynamic ranges.Without normalization, larger-valued features exert stronger influence on the classifier’s cost function.
- E. Feature Normalization: The normalization formula uses the original feature value together with the smallest original feature value to compute a new feature value.The supplied passages identify the roles of the new value, original value, and minimum original value.
III. PROPOSED SYSTEM
The proposed system segments a leaf image, extracts features, classifies the resulting vector with a trained PNN, and translates the output index into a plant name.
- III. PROPOSED SYSTEM: The Feature Extractor processes the input leaf before the PNN receives the extracted feature vector.The system schema places feature extraction before classification.
- III. PROPOSED SYSTEM: The PNN must be trained before classification and produces an index representing a plant, which the Plant Information Getter converts into the plant name.The system therefore maps extracted image features to a plant label through the classifier and index-translation stages.
- III. PROPOSED SYSTEM: Leaf segmentation builds a 20-bin intensity histogram, identifies two major peaks, and thresholds at the median of the smallest intervening bin.The two peaks represent the leaf and its background.
- III. PROPOSED SYSTEM: Morphological operations remove thresholding-induced holes, and an AND operation between the RGB and binary images obtains the leaf.This process is illustrated in Fig. 6.
- III. PROPOSED SYSTEM: PNN classification uses an input vector, training-derived weight vectors, class-level probability summation, and an output probability vector.Its pattern layer measures distances to training inputs, while the summation layer aggregates contributions for each class.
- III. PROPOSED SYSTEM: Under the PNN decision rule, an input is assigned to class j when p(x | w_j) exceeds p(x | w_i) for every i ≠ j.The conditional probability is calculated using the Parzen method and Gaussian-based density estimation.
IV. EXPERIMENTAL RESULTS
The proposed system was evaluated on the 32-plant Flavia dataset using separate training and testing samples. Combining shape, color without kurtosis, vein, and texture features produced the best reported accuracy, 93.75%.
- The Flavia dataset contains 32 kinds of plants and was used to test the proposed leaf-classification method.
- Forty plants per species trained the network, while 10 plants per species tested performance with a PNN smoothing factor of 0.05.
- Performance was calculated using the reported ratio of relevant images to total queries.The passage defines nr as the relevant-image count and nt as the total number of queries.
- 93.75% accuracy was achieved by combining shape, color without kurtosis, vein, and texture features.
- All feature types contributed importantly to performance except kurtosis.
V. CONCLUSIONS
The method combines shape, vein, color, and texture features with a PNN classifier and achieves 93.75% accuracy. The authors report that this is slightly better than the original work, while indicating that further improvement remains possible.
- The method incorporates shape and vein, color, and texture features and uses a PNN classifier.
- Fourier descriptors, slimness, roundness, and dispersion represent shape, while color moments represent mean, standard deviation, and skewness.
- Twelve texture features are extracted from lacunarity.
- 93.75% accuracy was obtained, slightly exceeding the original work’s reported 90,312% accuracy.
- The authors state that performance could still be improved and identify other features as future research.