چكيده به لاتين
Abstract
Due to unprecedented technological developments, the remote sensing field has been vastly expanded in terms of application recently, and in parallel, the performance has significantly improved. Object detection in remote sensing images faces more challenges because of more complex background information they contain than that of natural images. Remote sensing images offer information about the texture, shape, and structure of ground objects, and they can be used for precise object identification. Therefore, in our work, the used dataset has been augmented by three different augmentation methods which are rotation, flip and flip-rotation augmentation. We evaluated the results by using average precision measure (AP) as well as we compared our results with the results of previous works from the literature. We found that our proposed RetinaNet detects objects better than the methods used in previous works by at least %2.84 performance gain in mAP. Additionally, we noted that the worst results of our method are for rotation augmentation. By analyzing the results and making many tests we could fix the problem. For solving this problem, we used template matching. We selected the ship class as a special case for doing our tests. It was shown that the proposed template matching algorithm could solve the rotation augmentation problem to a high extent.
Keywords: Remote Sensing, Object Detection, Data Augmentation, RetinaNet.
