InaTechShips: A validation study of a novel ship dataset through deep learning-based classification and detection models for maritime applications
Abstract
The automatic detection and classification of ships in a maritime environment offer potential applications, ranging from calculating vessel traffic in a specific region to implementing security and defense systems. The consistency and size of datasets are important factors for the good performance of object detection models based on Convolutional Neural Networks (CNNs). The maritime vessel datasets currently available in the literature have certain limitations, such as the number of images, the use of very specific classes, and the lack of labels, making them unsuitable for specific applications. To address this issue, a public ship detection dataset called InaTechShips was created, comprising over 3 million images of maritime vessels, contributing to the state-of-the-art with accurately labeled images in different scenes and category variations. The obtained images were labeled automatically, representing a significant advancement over the manual method used by various authors. In this work, the performance of the new dataset is also evaluated through metrics such as Accuracy, Precision, Recall, F1-Score, mean Average Precision (mAP), Specificity, Error Rate (ERR), False Positive Rate (FPR), and False Negative Rate (FNR). In the validation scenario, various detection and classification models were used to evaluate the dataset's quality against models of different architectures. The models were trained with a portion of the database and showed promising results, as presented in the final chapters of this work. The codes necessary for downloading the database, automatic labeling, the generated label files, and trained models are all publicly available at (https://www.github.com/EduardoHT/InaTechShips/).