Unsupervised Learning Neural Network for Classification of Ship-Hull Fouling Conditions

We have developed algorithms to automate image classification and quantification of fouling conditions for ship hulls. Motivated by the fact that existing methods of quantifying fouling conditions and physical oceanographic conditions (e.g., coral reef population) have to rely on manual and visual labors by "experts", which are highly subjective and time consuming, we have developed algorithms to classify and quantify fouling conditions of ship hulls from images taken underwater. The algorithms include three parts: image production, feature extraction, and neural network classifiers. Raw images reflecting various fouling conditions were produced. Sub-sampling was conducted on the raw images and a total of 360 sub-image samples were generated. The 360 sub-image samples were divided into two datasets, one with 300 samples for training and the other 60 samples for testing of a self-organizing map (SOM) neural network. For each image, a total of 32 feature variables extracted from the statistical method of SGLDM were used for training the SOM neural network. Results show that with adequate training, the success rate for prediction can reach 100%.