Butterfly Image Identification Using Multilevel Thresholding Segmentation and Convolutional Neural Network Classification with Alexnet Architecture

Abstract

Lepidoptera is the name for the broad group of butterflies. The ecology depends heavily on butterflies, thus it is problematic that so little is known about their many kinds. Understanding butterflies is a crucial part of education since they are a natural occurrence and may be used as teaching tools. A total of 419 butterfly photos were utilized in the data. The dataset is first input, and then it undergoes preprocessing steps like segmentation, scaling, and RGB to grayscale conversion. CNN with AlexNet architecture is used to classify the preprocessed dataset's output. The outcomes of the classification stage of the Alexnet architecture are Flatten, Danse, and ReLu (Convolution, Batch Normalization, Max_Pooling). The output data is assessed following the completion of the Alexnet CNN training process. The data's ultimate classification is based on species. High-accuracy picture classification can be achieved using the model without segmentation, however, this cannot be achieved with multilevel threshold segmentation. According to the test findings, the multilevel threshold segmentation model only attains 62% accuracy, but the segmentation-free model gets 83% accuracy. The test results demonstrate that combining AlexNet architecture with multilevel thresholding segmentation resulted in a classification model that is less accurate in identifying different species of butterflies. By comparing these test results, it is possible to draw the conclusion that the multilevel threshold segmentation model performs less well at information classification than the model without segmentation.