Comparison of the Performance of Different Learning Algorithms in Leaf Feature Extraction and Recognition Using Convolution Neural Network

Abstract

Plant identification with computer systems has been developed with image processing tools and has helped researchers to identify unknown plant species with high accuracy. In this study, the leaves of five different plants were classified according to their shapes using deep learning. A database was created with leaf images of mint, echinacea, St. John's wort, melissa, and thyme plants. Images in this database were classified with a convolution neural network (CNN). For this classification, 70% training and 30% testing were randomly selected in the database. The parameters of the CNN layer consist of a set of 120x160x3$$ 120\times 160\times 3 $$ learnable filters. In the CNN, 10 3x3$$ 3\times 3 $$ kernel matrices with stride [1 1] were used. A rectified linear unit was chosen as the activation function. Maximum pooling was performed using a 2x2$$ 2\times 2 $$ filter with stride [2 2]. In this classification, five fully connected layers were created. Using CNN, the performance of different learning algorithms was compared. It was observed that CNN achieved more successful results than traditional attribute methods.