engleski

On detection of small objects from aerial images

The concept of small objects refers to those objects represented with a small number of pixels compared to overall number of pixels in the entire image. In addition to normal object detection challenges, small object detection is having additional challenges because it is hard to distinguish them from the background. The use of convolutional neural networks has made significant improvements in object detection performance. Various deep learning based approaches were presented by researchers in last decade. Some of them are having one stage such as SSD and YOLO and others consists of two stages (region proposal and classification stage) such as Fast R-CNN and Faster R-CNN. During last year, our research group proposed some new approaches in small object detection with application in detection of humans for search and rescue missions. These novel approaches include multimodel approach that combines two different convolutional neural network architectures in the region proposal stage as well as in the classification stage. Also, research related to size optimization of used neural network in order to provide real time on edge areal image processing. These and some other contributions will be discussed. Also, the ideas for the future improvement of current algorithms and detection results will be presented.

small object detection, aerial images, deep learning

nije evidentirano