Automatic Ship Classification Utilizing Bag of Deep Features

TL;DR

This paper introduces a novel ship classification method combining deep convolutional features with a bag of words approach, achieving over 91% accuracy and outperforming previous techniques.

Contribution

It presents a new approach that integrates deep features from pre-trained CNNs with the bag of words model for improved ship classification accuracy.

Findings

Achieved 91.8% classification accuracy.

Improved recognition performance by about 5%.

Utilized VGG models and SIFT for region proposals.

Abstract

Detection and classification of ships based on their silhouette profiles in natural imagery is an important undertaking in computer science. This problem can be viewed from a variety of perspectives, including security, traffic control, and even militarism. Therefore, in each of the aforementioned applications, specific processing is required. In this paper, by applying the "bag of words" (BoW), a new method is presented that its words are the features that are obtained using pre-trained models of deep convolutional networks. , Three VGG models are utilized which provide superior accuracy in identifying objects. The regions of the image that are selected as the initial proposals are derived from a greedy algorithm on the key points generated by the Scale Invariant Feature Transform (SIFT) method. Using the deep features in the BOW method provides a good improvement in the recognition and classification of ships. Eventually, we obtained an accuracy of 91.8% in the classification of the ships which shows the improvement of about 5% compared to previous methods.