Microscopic Image Based Classification Algorithms for few Herbal Plants Identification using Machine Learning Approaches

Abstract

Identification of herbal plant is of great interest in image processing and computer vision. In literature, many methods are presented to identify the herbal plant like leaf-based identification, chemical-based evaluation, physical evaluation and biological evaluation. Identification of herbal plant is more difficult and challenging when it is presented in powder form. This thesis presents microscopic image-based classification of a few herbal plants from its powder using various machine learning approaches. In this work, the cell characteristics of the herbal plants are studied. The dataset of the powder microscopic images of the three herbal plant is created in our laboratory using a Lawrence & Mayo microscope. It is found from the literature survey that various object can be uniquely represented by shape and texture based features which are further used for the object classification. In this thesis, shape and texture feature based novel methods for classification of herbal plants (Liquorice, Rhubarb and Dhatura) are proposed. The effectiveness of shape and texture feature methods are evaluated using different classifiers for classification. Three shape and five texture features are computed from the microscopic image dataset of the herbal plants. The effectiveness of the shape and texture based feature set and their combinations are investigated using Support Vector Machine (SVM), K-Nearest Neighbour (KNN) and Ensemble classifier. The effect of Speeded-Up Robust Features (SURF) is also analyzed using a different kernel of the SVM classifier. From the experiments highest classification accuracy of 99.9% is achieved when all the shape features with a combination of Gabor wavelet features are applied to quadratic SVM classifier. Finally, from the proposed algorithm, it is observed that the combination of selected shape and texture features work better for the classification of powder microscopic image of the herbal plant of Liquorice, Rhubarb and Dhatura. Convolutional Neural Networks (CNN) based classification is also carried out using pretrained CNN models like AlexNet, GoogLeNet and VGG for the herbal plants. Transfer learning approach is applied to train the pre-trained CNN models for our dataset. From the experiments, it is found that AlexNet-TL performs better and the results are compared with other pre-trained model GoogLeNet and VGG. It gives the highest accuracy of 99.49% with 40x dataset when the single resolution dataset is considered for training CNN. In the case of the combination of the two datasets of different resolutions, VGG16-TL achieves the highest accuracy of 99.96% with 4x and 40x dataset. The combination of two or more datasets of different resolution increases the training complexity in the CNN model. GoogLeNet-TL performs better when all three resolution of the dataset are combined. It achieves the highest accuracy of 99.03%. The fourth category of the class where the background of the image has been included with the dataset to train the pre-trained CNN model. In this category of dataset, AlexNet-TL gives 98.65% accuracy which is better than other pre-trained CNN model. It is observed from the simulation results, transfer learning from the pre-trained CNN model works better for automatic classification and highest classification accuracy of 99.96% is achieved in VGG16-TL for the combination of 4x and 40x dataset. When dataset of 4x, 10x and 40x is applied to the CNN for training, transfer learning work well with 40x dataset. It is due to the clear visibility of cell characteristics at 40x resolution. The features extracted from the microscopic image of the herbal plant using shape and texture are hand crafted features. It is very difficult to tune hand crafted features with a classifier for the herbal plant classification. In the CNN based classification of herbal plants, features are self-extracted from the input image dataset. CNN based classification is the good choice when input image dataset is complex and difficult to tune the hand crafted features. It takes the more training time to train CNN model. This problem can be overcome by using the Graphic Processor Unit (GPU) during the training of CNN.