MorphoNet: A Novel Bivalve Images Classification Framework with Convolutional Neural Network

Abstract

The bivalves' morphometric analysis of the freshwater shell characteristics is based on the shell size, shape, tooth, scars, and texture. We experimented and compared the accuracies of the following popular convolutional neural network architectures: ResNeSt, MobileNet, VGG16, Transfer Learning, and EfficientNet, whose model trainings are based on the bivalve image dataset obtained from a biology laboratory. The MobileNet model that gives the highest accuracy rate by 72% is selected to be a classification model of our framework named MorphoNet. We also applied the YOLO4 object detection in the MorphoNet to detect the teeth and scars on the bivalve image. The framework can identify the bivalve class labels and detect the interesting features on the bivalve images automatically. It is an alternative tool to help the biologists in a preliminary class label identification and support the land-marking creation and morphometric analysis instead of doing it by hand.