Gabor Feature Representation and Deep Convolution Neural Network for Marine Vessel Classification

Abstract

The Vessel Surveillance System (VSS), a crucial tool for fisheries monitoring, controlling, and surveillance, has been required to use for the reservation of the current depressed state of the world's fisheries by fisheries management agencies. An important issue in the vessel surveillance system is the classification of vessels. However, several factors, such as lighting, congestion, and sea state, will affect the vessel's appearance, making it more difficult to classify vessels. There are two main methods for conventional classifications of vessels: the traditional-based- characteristics method and the convolutional neural networks-used method. In this paper, we combine Gabor feature representation (GFR) and deep convolution neural network (DCNN) to classify vessels. Gabor filters in different directions and ratios are used to extract vessel characteristics to create a new image of vessels, which is DCNN's input. The visible and infrared spectrums (VAIS) dataset, the world's first publicly available dataset for paired infrared and visible vessel images, was used to validate the proposed method (GFR-DCNN). The numerical results showed that GFR-DCNN is more accurate than other methods.