Butterfly Image Identification Using Multilevel Thresholding Segmentasi and Convolution Neural Network Classification with Alexnet Architecture

Abstract

Butterflies form a large group called Lepidoptera. Butterflies play an important role in the ecosystem so the lack of knowledge about butterfly species is a problem. Since butterflies are a natural phenomenon and can serve as an educational tool, knowledge about butterflies is an important component of education. The data used totaled 419 butterfly images which were divided into two, namely training data and testing data. First, the dataset is input, and then the dataset is preprocessed such as resizing, converting RGB to grayscale, and segmentation. The output of the preprocessing dataset is classified using CNN with AlexNet architecture. The results of the Alexnet architecture classification stage include ReLu (Convolution, Batch Normalisation, Max Pooling), Flatten, and Danse. After the Alexnet CNN training process is complete, the output data is evaluated using the calculation of Accuracy, Precision, and recall. The final result of the data is classified according to the species, the model without segmentation is able to classify the image with high accuracy, while using multilevel threshold segmentation cannot classify the image with high accuracy. The test results show that the model without segmentation has 83% accuracy, while the model with multilevel threshold segmentation only achieves 62% accuracy. The test results show that the combination of multilevel thresholding segmentation and AlexNet architecture creates a classification model that is less accurate in recognizing butterfly species. Comparing these test results, it can be concluded that the model without segmentation tends to be better at classifying information than the model using multilevel threshold segmentation.