Review Edge and Texture Fusion for Plant Leaf Classification

   


 

1. summary

Plants are an important resource in life, which can be developed in the fields of industry, medicine and food. classification plants with automated digital technique, can be done simply by extraction features of the leaves. In this paper, the classification of plants is based forms and texture. Based on the theory of plant taxonomy, plants can be classified based on the shape of the leaves and flowers. However, in general, the classification is based on leaf analysis. Classification based on the shape of the flower is difficult due to the structure of interest is 3D and can only be obtained when the flowers bloom.

Feature extraction used in this paper is the Sobel edge detection and Gabor filters. In edge detection, there are three stages: filtering, enhancement, and detection. Sobel edge detection find the approximate absolute gradient magnitude at each point to detect edges. In Gabor filters have the ability to do the multi-resolution decomposition, and are ideal for segmenting texture in the spatial and spatial-frequency domains. Then the method of feature extraction are tested on classification and regression tree (CART) classifier and the Radial Basis Function (RBF) classifier. Variation of the RBF NN is better in terms of interpolation, by taking input non-linear and linear output. Cost function is usually square error is assumed to be used to evaluate the suitability of parameters. Activation function using a Gaussian function by the equation:
Where X = the input feature vector, L = number of hidden units, uj = average, and Σj = convariance matrix of the Gaussian function.

This paper shows the degree of accuracy RBF classification higher than in CART. The level of accuracy of classification in RBF squared error of 85.93 with 56.2. While the CART rate of accuracy classification of premises 79.26 67.38 squared error. So that RBF has a level squared error is smaller than CART.

2) Excess

  • Gabor filters have the ability to perform multi-resolution decomposition
  • Gabor filters are ideal for texture segmentation because it requires the simultaneous measurement the spatial and spatial-frequency domains.
  • Better in terms of RBF interpolation, taking a non-linear input and provide linear output, thus helping pemetakan complex models.
  • Sobel edge detector find the approximate absolute gradient magnitude at each point to detect edges.

3) Weakness

  • RBF output is limited to the interval (0,1)

4) Suggestions

  • Improving the classification accuracy of the proposed feature reduction technique.


paper