Bilişim Enstitüsü, Uydu Haberleşme ve Uzaktan Algılama Programı, İTÜ
Automatic airplane detection using deep learning techniques and very high-resolution satellite images
Bakary Traore; Elif Sertel, 2020
Abstract: Object detection in high resolution remote sensing (RS) images is challenging problem in remote sensing imagery processing, especially for civil and military application, due to the complex object background and complex scenes. Many researches were conducted and are also under process in the field of remote sensing using deep learning (DL) techniques, a state-of-the-art technique, for automatic object detection. Some of them successes as well as others have failed. The successful researches have also faced some problems: High training time; low detection performance; high detection time; poor localization accuracy. Also, in their studies, different models lead to different results on a same given dataset. The main concern in this thesis is to figure out the reasons of these problems by analyzing different components of some of the available models. In this research, we proposed airplane detection approach using different deep learning techniques applied to high-resolution satellite images. Two different well-known DL techniques such as Single Shot Multibox Detector (SSD) and Faster Region-based Convolutional Network (Faster RCNN) were used for our airplane detection. These techniques are the most used deep learning techniques for object detection. We present fifteen models with seven models based on SSD and eight based on Faster R-CNN. They were pretrained using Microsoft Common Objects in Context (MS COCO) dataset. Each model was settled with specific structure and hyper-parameters over time in order to determine the importance of the model architectures and hyper-parameters in object detection to overcome the above-mentioned problems. During the experiments, we encountered overfitting problem with some Faster R-CNN models, we overcame the problem by using regularizers with different weight values. The models were trained with a well-built dataset composing of small, medium, and large-scale airplane and non-airplane objects. Moreover, our dataset is a mix of images of high-resolution satellite images from different sources such as Northwestern Polytechnical University-REmote Sensing Image Scene Classification (WPU-RESISC45) dataset, WHURS19 dataset, Aerial Image Dataset (AID), and Istanbul Technical University-Center of Satellite Communication and Remote Sensing (ITU-CSCRS) dataset, used in several object detection applications. We used images from these datasets because there are composed of images with different resolutions, different scenes with different backgrounds, from different sensors and earth’s regions. Also, the datasets were used in many scene classification and object detection projects and they produced state-of-the-art results. All images from these sources are very high-resolution satellite images and mainly created for scene classification and object detection. Our dataset, with this variety of sources, is composed of 1402 images in total with 7 701 airplane objects in 1167 images other images contain non-airplane objects such as Jet Plane, High Building, Residence, Storage Tank, and Pool. We used 80% of the dataset for training with 6 287 plane objects in 1 119 images and the remaining for testing which includes 1 414 plane objects in 283 images. We achieved airplane detection with fast training, accurate detection, and detecting targets regardless of their background and scene configuration. At the end, the models were analyzed and compared in terms of their performances on high-resolution satellite images, where we found out that the model architectures and hyper-parameters, for instance feature extractor and learning rate and many others, have a significant role in model training and the accomplishment of a desirable object detection result. It is our hope that this research will be beneficial for the researchers to avoid above mentioned problems in their future studies and applications.
Tez No: 615568