Performance Evaluation Of Image Recognition Algorithms On Marine Vessels And Optimum Parameter Selection

Abstract

Today, with the developing sensor technology, image processing models and deep neural network methods, there are significant developments in the field of autonomous driving and also various studies are carried out in this direction both in the private sector and in academia. On the other hand studies on safe driving of autonomous vehicles are still very limited. Mostly the studies have been conducted for land vehicles, and the data sets for the operation of artificial intelligence models were created in this context. In this study, the algorithms for autonomous driving were tested using the original data set created from objects on the sea in order to optimize the navigation of sea vehicles on the sea. Image processing methods have recently gained great importance in terms of recognizing vehicles on the sea and providing autonomous driving. In this study, a high-resolution and wide-ranging original data set consisting of 44965 objects was created to identify objects on the sea. With this data set, analysis and optimizations were made with image processing technology for the recognition and classification of objects, and the best model was tried to be determined among the models. It is aimed to detect and classify objects on the sea surface from a long distance (1000m+), to create safe use for sea vehicles and to provide decision support. In order for the created data set to be successfully identified in real-time, the data set was divided into six classes. As a result of the classification process, data labeling was performed according to these classes which are, Cargo_Ship, Tanker_Ship, RoRo/Ferry/Passenger, Boats, Tug_Boats, Speciality_Vessels. The created data set was tested with the most common real-time recognition models, SSD, Faster R-CNN, EfficientDet algorithms under the TensorFlow library. Results were obtained according to six different output parameter values, AP-50, AP-75, Av. Recall, F1-50, F1-75 and L/TL, on the models. According to the obtained results, SSD Mobilnet v1 was determined as the most successful algorithm.