Volume: 19 Issue-01 (January-June) 2024

---

TRANSFER LEARNING IN CONJUNCTION WITH MULTI-OBJECT DETECTION USING YOLO RCNN

ABSTRACT

Processing images from satellite photography is a significant difficulty. Finding things in a satellite picture is a crucial job in this field. Since there has been a lot of study done on machine learning-based image processing, machine learning techniques may be used for this. Image processing tasks may be carried out via a multitude of machine learning-based supervised and unsupervised techniques. This research evaluated object identification systems based on machine learning using satellite photos. Finding one supervised and one unsupervised method to apply and compare across a created dataset was the aim of the study. In order to tackle these obstacles, our research used transfer learning in conjunction with multi-object identification deep learning algorithms using remotely sensed satellite data obtained on a diverse terrain. The models in the research were assessed using a fresh dataset of varied characteristics with five item classes that were gathered from Google Earth Engine at different places in the southern South African province of KwaZulu-Natal. The items in the dataset photos varied in terms of size and resolution. Using our recently developed dataset, five object identification techniques based on YOLO R-CNN architectures were examined via tests. A survey of the literature was done to determine the best algorithms for object recognition. For supervised learning, support vector machines as well as k-means were used, respectively, based on the literature review. To put these algorithms into practice, an experiment was conducted. A dataset of items from satellite pictures was produced specifically for the project. Both silhouette score analysis and confusion matrix analysis were used to assess the experiment's outcomes.The findings indicated that YOLOv8 had the quickest detection speed of 0.2 ms and the greatest detection accuracy of more than 90% for situations including vegetation and swimming pools. The study's findings indicate that, when it comes to object detection on satellite photos, support vector machines are more useful than k-means clustering.

FULL TEXT

PDF