Plant Leaf Classification Free Essay Example

To install StudyMoose App tap and then “Add to Home Screen”

Abstract

In the plant leaf classification, leaf representation with traditionally handcrafted features is difficult to reveal its complex shape and texture. In this paper, we proposed a novel plant leaf classification method based on convolutional neural network (CNN) due to its powerful capability of feature extraction and classification. In our method, a ten-layer CNN was constructed for plant leaf classification. In order to improve the classification, sample augment for leaf was applied to the images to enlarge the database. The visualization was utilized for analyzing the factors influencing the accuracy rate.

Don't use plagiarized sources. Get your custom essay on

“ Plant Leaf Classification ”

Get custom paper

NEW! smart matching with writer

The experimental results on leaf database Flavia with 4,800 leaf images and 32 kinds of leaf showed that the proposed method achieved a high overall accuracy with 87.92 percent.

Keywords: deep learning; convolutional neural network; leaf classfication; feature extraction I.

Introduction

Leaf can be collected conveniently and preserved for a long time. The great difference among the majority of diverse species is apparent and the shape of the same species is similar. For example, the leaves of rosaceous plant generally are oval while the poaceae plant generally are lanceolate.

Hence, leaf is one of the main reference points for leaf classification. Generally, the extraction of leaf features, such as traditional shape feature, texture feature and color feature, is important for leaf classification. Based on the extracted feature, then machine learning or pattern matching is used to classify. Yang et al. represented the leaf contour with the ordered sequence and used the amplitude-frequency feature to recognize the leaf with high accuracy on ICL library.

Priya et al used support vector machine classification via extracting five fundamental features from Digital Morphological Features.

In addition, the leaf edge variation was considered as the feature to represent the leaf. In the combination of texture and shape feature can represent better the leaf than single feature. Then two neural networks were used to classify the leaf. Above methods mainly used manual feature extraction to represent the leaf and then various machine learning methods were used to classification. In recent years, deep learning methods such as convolutional neural networks (CNN) and deep belief networks (DBN)show the excellent performance in object detection and classification. Sue Han Lee et al. classified 44 different plant species with CNN based on the AlexNet. Two datasets including the whole leaf images (D1) and patches of the leaf images (D2) were trained with CNN. The accuracy rate of D1 was 97.7 percent and D2 was 99.6 percent. Ferrira et al proposed a method to classify the weeds among grass and broadleaf using the AlexNet. The work achieved above 98% accuracy in the detection of broadleaf and grass weeds in relation to soil and soybean. The accuracy average between all images above 99 percent. Konstantinos P. Ferentinos proposed a plant disease detection and diagnosis method using CNN. He used an open database of 87,848 images, containing 25 different plants. This method reached a 99.53 percent success rate. At present, some papers proposed that AlexNet or VGG is used in classification to identify images with complex content . At this point, the large input size is required and the neural network will consume more time and internal storage.

In this paper, a ten-layer neural network is presented because the leaf images have simple content which also can reach an expected result and decrease the training time and storage. While the over-fitting is still a problem to be solved for deep learning. The five traditional methods and three deep learning methods are introduced in the above. The accuracy rates of deep learning methods are higher than the feature extraction methods. And the traditional method depends on the character of the database, so it may not be generalizable. The deep learning performs better on image classification. In this paper, we proposed an approach for designing a CNN to distinguish between 32 plant species. The main opinion of a CNN was to build a hierarchy network. Moreover, we analyzed the experiment result and conclude the factors influencing the accuracy rate. In section , we introduced the data materials and the approach for the leaf classification. In section we analyzed the recognition results from two aspects, first, we compared the CNN with traditional method, second we analyzed the misclassifications. In section , we concluded this paper.