Plankton recognition using similarity learning

Abstract

Several automated classification methods for plankton images have been developed. These methods typically require an explicit description of features, data augmentation, and are not suitable for classes with a few example images. In recent years, deep metric learning has become a popular technique for many applications. This thesis focuses on classifying plankton images using a convolutional neural network (CNN) based on metric learning. The basic idea is to develop a model trained with a Triplet loss function using CNN architecture for feature extraction. The training model decreases the distance between similar objects and increases the distance between dissimilar objects. The classification is done by searching the image from the set of examples with minimum distance to the query image. The result shows that the accuracy of most plankton classes is good with similarity learning even with a few example images in an unbalanced dataset without data augmentation. However, some plankton images were not classified correctly. Further investigating of architectures and similarity classifier is needed to improve the accuracy.