Development of artificial intelligence based methods for underwater litter detection

Marine pollution is one of the biggest environmental problems in the world. It is estimated that between 100 and 150 million tons of plastic litter are floating in the oceans and seas today. These numbers are increasing every year. Plastic wastes left to the sea can continue their existence in the seas for centuries. Thus, it adversely affects the lives of living things, especially sea creatures. In this thesis, artificial intelligence-based methods have been developed for litter detection from underwater images. Within the scope of the thesis, image processing, machine learning and deep learning based methods were applied on underwater images. A new data set was collected for the application of the proposed method. In addition, data sets commonly used in the literature were used and collected in different images from internet sources. The proposed methods have been applied for many data sets and experimental results are given. Within the scope of the thesis, scientific contributions were made in three main points. First, the dataset was collected using the GLADIUS MINI model wired underwater robot. On this data set, litter is labeled according to its types. Using the YOLO-V4 algorithm, 88.7% success was achieved for the 80:20 training test data. Secondly, a method based on image processing and K Nearest Neighbor Algorithm (KNN) is proposed for underwater image classification. The Trash-ICRA19 dataset was used to test the proposed method and compare it with the results in the literature. The dataset cropping process was applied and a dataset consisting of 11060 images in total was obtained. These images were converted to 200×200 pixels using preprocessing. By applying the Directional Gradient Histogram (HOG) algorithm, 11060×900 feature vectors were obtained. The results obtained have 97.78% accuracy when the KNN classifier is used in this method. Third, a projector deep feature extraction based litter image classification model using underwater images is proposed. A hybrid dataset was collected to test this method with a large number of images. Feature extraction was done on the image using the projector ResNet101. Extracted features were combined and the most weighted features were selected using Neighborhood Component Analysis (NCA). Selected features were classified with the KNN algorithm. 99.35% accuracy was calculated with the proposed method. As a result, methods for detecting underwater litter are proposed and performance results are presented within the scope of this master thesis. The studies carried out within the scope of the thesis were supported by the 2210/C Domestic Priority Areas Graduate Scholarship Program with project number 1649B022204832 (TÜBİTAK) and by the FÜBAP graduate thesis research projects with the number TEKF.22.01.