Classification of leaf images using convoluti̇onal neural network

We do not have information about many plants around us, we act with the information among the people about their benefits and harms. In such cases, the wrong use of herbs known as medicinal, although toxic, causes many problems that can lead to death. In such cases, expertise is required to identify plants. This problem is a deficiency that can be eliminated with developing technology without requiring expertise. For this reason, it is aimed to create a ready system using artificial intelligence techniques on leaf images. Nowadays, where artificial intelligence technologies are at the forefront, many methods are used in this field. One of these methods is deep learning. We can say that deep learning methods are used in most areas. Areas such as face recognition, voice recognition, autonomous vehicles, cyber threat analysis, healthcare. While using this method, it serves many purposes such as definition, classification, and prediction. Convolutional neural network, one of the deep learning algorithms, learns the features with the data it receives as input and obtains the desired result with the help of these features. It is seen in the literature that the data obtained by using different feature extraction methods are used with different classification methods. There are approaches such as using the features extracted from the created convolutional neural network models or pre-trained models with different classification methods. In this thesis, data sets consisting of different numbers of leaf images were used. A single large data set consisting of 270 different species including all images was created and a new convolutional neural network model specific to this data set was trained and the results were examined. Improvement studies have been carried out by examining the effect of changes such as color or gray image difference, effect of background color, different image size and filter size. The structure of the model created with the data obtained as a result of these studies and the status of the data to be used in the training of this model were determined. Leaf classification has been made using the convolutional neural network model. Also, using this model, feature transformation has been performed with the approach of feature learning and its performance has been investigated with different classification methods. As a result of the size of the data set, it was ensured that pre-trained models that do not need training due to the need for powerful equipment for using and the results were compared. For this purpose, features were extracted from 4 different pre-trained models and classified with the same classification methods. By evaluating the data obtained as a result of the classification process, an approach to achieve the best result in hardware with low processing power was determined and 89% training, 86% test performance were achieved for the data set with 270 species.