Hide/Show Apps

Hyperspectral data classification via capsule networks

Soyak, Elmas
In this thesis, a novel deep architecture capsule networks are investigated for hyperspectral data classification purposes. Even though this algorithm resembles convolutional neural networks (CNN), which is one of the most successful methods in classification, capsule networks have been developed to overcome the limitations of it. CNN applies convolution operation to extract features in the samples and uses these features to classify them. However, it fails to measure the relationship between these features. Moreover, pooling operation that is used in CNN to reduce the number of parameters results in loss of position information and thus decreases the success of classifier. With the novelties proposed in capsule networks, it is intended to resolve the shortcomings of CNN mentioned above. Instantiation parameters such as position, orientation, scale of each feature are kept in a capsule, and both the presence of the relevant feature and the instantiation parameters of the feature are utilized in the classification step. In the experiments performed on hyperspectral data, the efficiency of capsule networks is evaluated by using different number and structure of training samples. A CNN algorithm with a similar structure is constructed and compared with capsule networks. Although the presented method yields successful results, it has been observed that iteration is exhausting in terms of memory and processing time.