Original papers
Rotation invariant wavelet descriptors, a new set of features to enhance plant leaves classification

https://doi.org/10.1016/j.compag.2017.05.031Get rights and content

Highlights

Abstract

Automatic plant leaf recognition can play an important role in plant classification due to leaf’s availability, stable features and good potential to discriminate different kinds of species. Amongst many leaf features like leaf venation, margin, texture and lamina, leaf shape is the most important one due to its better discriminative power and ease of analysis. One of the most common leaf shape descriptors is Elliptic Fourier Descriptor (EFD). In this paper a new shape descriptor is introduced as “Rotation Invariant Wavelet Descriptor” (RIWD). The performance of RIWD is compared with IEFD using Flavia dataset. MLP neural network is used as the classifier in this work. Results analysis shows better performance of the proposed feature in classification accuracy. Furthermore, an optimum feature vector is constructed using a set of textural and morphological features and the RIWD that reached 97.5% classification accuracy with low computational cost in comparison with many reported results in Flavia dataset.

Introduction

Plants are substantial for the survival of human society. Recently many species of plants are at the risk of extermination. Thus it is very essential to set up a dataset for plant variety protection (Wu et al., 2007). It is very beneficial to distinguish which plants are eatable and useful, and which ones produce irritation or even death. There are about 3 million species of plants that have been named and classified on earth (Harish et al., 2013).

Traditionally, professional taxonomists are trained to identify species and their relationships. They can examine specimens and assign taxonomic labels to them. However, a problem known as the “taxonomic impediment” (de Carvalho et al., 2007), is caused by deficiency of such experts. Moreover, an expert on one species or family may be unfamiliar with another. This has resulted in an increasing interest in automating the process of identification of plant species (Cope et al., 2012).

Automatic plant identification is a challenging task. Plant organs like flowers, fruits and leaves can be used for identification. Flowers and fruits are often only available for a few days or weeks of a year. Therefore, leaf recognition plays an important role in plant classification due to its availability, stable features and good potential to discriminate different kinds of species (Singh et al., 2010).

An image processing approach that leads to automatic segmentation of plant leaves can play an important role in such a task. Even some mobile applications are developed to do this (Grand-Brochier et al., 2015). Actually, leaf shape, venation, margin, texture and lamina (surface) are common leaf features involved in several applications (Cope et al., 2012). However, some researchers used part of those features only (Kadir et al., 2013).

Amongst the abovementioned leaf features, leaf shape is the most important one due to its better discriminative power and ease of analysis. Wu et al. (2007) used 5 basic geometric and 12 morphological features as leaf shape descriptors, PCA analysis and PNN as the classifier for leaf recognition. They also introduced a well-developed dataset called Flavia.

Flavia dataset was used by Singh et al. (2010) that did a research to compare Wu’s algorithm with other methods including Support Vector Machine (SVM) and Fourier moments (Kadir et al., 2013).

Zernike moments, a set of shape descriptors, were implemented by Kulkarni et al. (2013). In this method, the extracted leaf features were combined with the Zernike moments. The focal improvements were based on the feature extraction techniques that included Zernike moments and the dual stage learning algorithm for training the classifier using Radial Basis Function neural network (Harish et al., 2013).

Wong et al. (2007) used Fourier descriptor as a similarity measure and support vector machine as pattern classifier for a large scale dataset. Yadav et al. (2008) reported retrieval and classification of various shapes using generic Fourier description.

Prasad et al. (2016) proposed an RST-invariant shape profile transform called Angle View Projection (AVP) that is used in leaf recognition application. Discrete Cosine Transform (DCT) is then applied to the resultant signal to maintain the shape property of leaf while reducing its dimension using Principal Component Analysis (PCA). They reported promising results in Flavia dataset.

Apleaf is an android-based plant leaf identification system (Zhao et al., 2015). The feature extraction phase of this system uses pyramid histograms of oriented gradients (PHOG), HSV color features and their statistical moments and wavelet coefficients. The system was tested using ImageCLEF2012 dataset which contains 126 tree species from the French Mediterranean area.

During an international competition called LifeCLEF plant challenge, 7 research groups submitted their results for plant recognition on PlantCLEF 2015 dataset. This dataset includes 113,205 images from 41,794 observations of 1000 species of trees, herbs and ferns living in Western European regions. The detailed description of each submitted method and results can be found in Joly et al. (2015).

One of the most common leaf shape descriptors is Elliptic Fourier Descriptor (EFD). Mancuso (2015) used Elliptic Fourier Analysis and artificial neural networks for grapevine genotypes identification. Neto et al. (2006) used Elliptic Fourier (EF) and discriminant analysis to identify 4 plant species, based on leaf shape. Invariant Elliptic Fourier Descriptor (IEFD) is developed for cereal grains classification that is independent of size and rotation of the shape contour (Mebatsion et al., 2012b). The variation of grain types is also evaluated using IEFD and PCA (Mebatsion et al., 2012a). Moreover, a dozen of research works has been reviewed by Cope et al. (2012) that used variations of EFD for plant leaf identification.

In this paper a new shape descriptor is introduced as “Rotation Invariant Wavelet Descriptor” (RIWD). The performance of RIWD is compared with IEFD using Flavia (Wu et al., 2007) dataset. Furthermore, an optimum feature vector is constructed using a set of textural and morphological features in combination with the mentioned descriptors to reach the best classification rate in Flavia dataset. MLP neural network is used as the classifier in this work.

Section snippets

Dataset description

The Flavia dataset contains 32 different species, a total of 1907 leaves. A full description of the species names and the sample numbers for each one can be found in Wu et al. (2007). Table 1 shows a few samples of the dataset.

Preprocessing

The images from dataset are converted to binary and preprocessed using morphological operations to eliminate some border errors. However for color features, the original images are used. The leaf contour is extracted from the preprocessed image, as shown in Fig. 1.

Flavia

Comparison of RIWD using different wavelet types with IEFD results

The proposed RIWD is used to classify the leaf images of Flavia dataset in this paper. The performance of RIWD is compared with IEFD. This performance is calculated using the classification accuracy which is given in Eq. (41).Classificationaccuracy=TP+TNTP+TN+FP+FNwhere TP, TN, FP and FN are True Positive, True Negative, False Positive and False Negative number of classified samples, respectively (Powers, 2011, Fawcett, 2006). The overall performance for each method is calculated from the mean

Conclusion and future work

The main purpose of this study was to devise a comprehensive approach that enhances leaf image classification accuracy. This aim is achieved by a new feature (RIWD) that is proposed based on Elliptic Descriptors. Previous Elliptic Descriptors used the Fourier coefficients to represent the shape of a leaf contour. However, RIWD uses the wavelet coefficients, instead. Using the popular Flavia dataset, the results comparison shows an improvement in the performance of the proposed shape descriptor.

Acknowledgements

The authors would like to thank Dr. Ali Gholami Rudi for his help on computational complexity calculations of this work. Also, they would like to thank the anonymous reviewer and editor for the useful and constructive comments which have improved the article.

References (30)

  • W.-T. Wong et al.

    Shape-based image retrieval using support vector machines, Fourier descriptors and self-organizing maps

    Inf. Sci.

    (2007)
  • R.B. Yadav et al.

    Retrieval and classification of objects using generic Fourier, Legendre moment, and wavelet Zernike moment descriptors and recognition using joint transform correlator

    Opt. Laser Technol.

    (2008)
  • Z.-Q. Zhao et al.

    ApLeaf: an efficient android-based plant leaf identification system

    Neurocomputing

    (2015)
  • T. Adamek et al.

    A multiscale representation method for nonrigid shapes with a single closed contour

    IEEE Trans. Circuits Syst. Video Technol.

    (2004)
  • M.R. de Carvalho et al.

    Taxonomic impediment or impediment to taxonomy? A commentary on systematics and the cybertaxonomic-automation paradigm

    Evol. Biol.

    (2007)
  • Cited by (32)

    • Leaf classification on Flavia dataset: A detailed review

      2023, Sustainable Computing: Informatics and Systems
    • Invariant leaf image recognition with histogram of Gaussian convolution vectors

      2020, Computers and Electronics in Agriculture
      Citation Excerpt :

      The classification accuracies of the proposed methods together with those of the benchmarks are summarized in Table 6. It can be seen that as a pure shape descriptor, the proposed HoGCV achieves a classification accuracy of 96.56% which is higher than the first two handcrafted descriptors, Morphological features (Wu et al., 2007), RIWD (Yousefi et al., 2017). Although the last two handcrafted descriptors, RIWD + morphological and textural features (Yousefi et al., 2017) and Shape + texture + venation features (Saleem et al., 2019), are 0.94% and 2.19% higher than the proposed HoGCV, they are not pure shape descriptors and incorporate other features such as texture and venation for description.

    • Maximum mutual information and Tsallis entropy for unsupervised segmentation of tree leaves in natural scenes

      2019, Computers and Electronics in Agriculture
      Citation Excerpt :

      To distinguish between different species, computer vision techniques can help to extract more characteristics from the leaves than traditional methods. Leaf segmentation from images under controlled conditions like white background is almost easy (Yousefi et al., 2017), but the automatic segmentation of the leaves from natural background scenes, which consists unconstrained image acquisition is a much more challenging task. Recent works on leaf image segmentation are still time-consuming or require special conditions.

    • Analysis of transfer learning for deep neural network based plant classification models

      2019, Computers and Electronics in Agriculture
      Citation Excerpt :

      The proposed method was tested with Flavia and Swedish Leaf datasets. Yousefi et al. (2017) used view rotation invariant features which are extracted from the Fast Fourier Transform and Discrete Wavelet Transform. Yu et al. (2016) developed a method utilizing the leaf contour and venation.

    • CCDF: Automatic system for segmentation and recognition of fruit crops diseases based on correlation coefficient and deep CNN features

      2018, Computers and Electronics in Agriculture
      Citation Excerpt :

      Feature extraction plays an important object identification technique in machine learning domain (Akram et al., 2018; Sharif et al., 2018b; Iqbal et al., 2018). In literature, several features have been investigated for object identification purposes such as shape (Zhang et al., 2017; Khan et al., 2018), texture (Zhang et al., 2017; Majdar and Ghassemian, 2017), color (Koviera and Bajla, 2013) and geometric features (Yousefi et al., 2017; Sharif et al., 2018c). Each feature has advantages in its specific context.

    View all citing articles on Scopus
    View full text