Sammendrag
Planktonic species are of great importance in the marine ecosystem, standing for about half of the earth's primary production and being a fundamental part of the marine food chain. Yet, our understanding of these vital creatures and the consequences of small changes in their habitats, creating profound shifts in planktonic dispersion and abundance, is still limited. Recent development and innovation within autonomous underwater vehicles (AUV) makes it possible to utilize AUVs for in-situ identification and classification of plankton taxa resulting in faster and more reliable calculations of its distributions than any existing sampling and classification platform. For accurate image classification, the system is heavily dependent on accurate and attested machine learning techniques. Currently, deep convolutional neural networks have proven especially effective for this task. Yet, the success is fixed to supervised learning, which requires an extensive amount of labeled training data. Such methods thus require a comprehensive and time-consuming labeling effort.
The scope of this work is to make a deep learning framework for plankton classification training on images that contain no ground truth labels. This work extends the work of the specialization project by proposing new feature extraction methods using state of the art unsupervised training schemes. These models can then be used to extract features that can improve a separate clustering algorithm. For comparison to the specialization project, the models are tested over existing planktonic data sets. The most successful methods are then adapted onto the image data acquired from the AUV missions in the Trondheim fjord.
Three methods for improved feature learning, DeepCluster [15], a generative adversarial network (GAN) model and a rotation-invariant autoencoder were selected for this thesis.
The former two were chosen as they represent some promising new directions within the unsupervised deep learning and have shown recent success at learning essential features on large image data sets. The rotation invariant autoencoder is an extension of a conventional autoencoder, which is more robust to similar class objects with different rotations. The proposed methods were then used as feature extractors to significantly improve different clustering algorithms in regards to classification performance over planktonic data. The chosen methods demonstrate some of the possibilities within the unsupervised domain, but the gap towards supervised learning is still significant.
Vis fullstendig beskrivelse