Comparison of MobileNet and VGG16 CNN Architectures for Web-based Starfish Species Identification System

Abstract

Bunaken Marine Park (BMP) is famous for its rich marine biodiversity. BMP is an asset for the marine tourism industry of the Manado city government, and the North Sulawesi Province of Indonesia needs to be strengthened. This research aims to build a web-based intelligent system using a convolutional neural network (CNN) to identify starfish species to initiate developing a media center marine biota identification system of BMP. Two CNN architectures, namely MobileNet and VGG16, were conducted to produce identification models. The first stage carried out a training process on 1800 starfish image data and then evaluated using the 5-fold cross-validation technique. Validation results show that MobileNet is superior to the VGG16 architecture by achieving validation accuracy of 100% for each fold while VGG16 produces validation accuracy in the range of 94% to 100%. On the other hand, in the second stage of model testing, it was found that VGG16 worked better than MobileNet in identifying 200 new data. The Best Model produced by VGG16 achieved testing accuracy of 100% while MobileNet produced 99.5%. However, stability analysis of the identification models produced by both architectures shows that MobileNet has relatively small loss values ranging from 0.00069325 to 0.00214802 as well as smaller standard deviation values of 0.27 compared to 0.61 produced by VGG16. These findings indicate MobileNet is more stable in carrying out identification work compared to VGG16 of, thus the best model provided by MobileNet is taken to deploy in the web platform which is created using the Python flask framework. The proposed system can be used to strengthen the marine tourism industry as a media center of educational marine biota using deep learning approaches.