Search

Due to the complexity of object information and optical conditions of high-resolution aerial imagery, it is difficult to obtain fine semantic segmentation performance. Although various deep neural network structures have been proposed to improve segmentation accuracy, there is still room for improving accuracy by making full use of multiscale features and integrating these single weak classifiers into a strong classifier. In this paper, we use a reduced SegNet network to realize the end-to-end classification of high-resolution aerial images. In addition, to use multiscale information, we present the R-SegUnet which combines the feature information of each convolution block in the reduced SegNet encoding network with the feature information of the corresponding convolution block in the decoding network. Furthermore, considering that the surface features in high-resolution aerial images are very complex, we investigate a 6to2_Net that converts the six-classification model into six binary-classification models for the recognition effect on small objects. Finally, we ensemble the above three different models to get the segmentation results. Experiment results on ISPRS Potsdam benchmark dataset show that our algorithm is state-of-the-art method. We also analyze the inference performance of our models on a variety of parallel computing devices.

ABSTRACTObject detection has played an important role in the remote sensing (RS) image understanding domain, but it still suffers from challenges, including the complexity of the image background, the variety of objects and differences of target scales. Although many studies applying deep learning methods to object detection have been presented, effective methods of multiscale object detection in RS images are still lacking. In this paper, we propose a novel object detection approach based on the Multiscale-SELU-DenseNet (MSE-DenseNet) and the dynamic anchor assignment (DAA) strategy. We fuse a highly modified DenseNet and a feature pyramid network into the MSE-DenseNet to extract multi-level feature descriptions of RS images, and then we use them to detect objects on various scales. To further improve the detection performance, we propose a DAA strategy for assigning prior anchor boxes. The experimental results show that our proposed method outperforms state-of-the-art detection methods on the VHR-10 dataset and maintains a real-time detection speed.

A ranking support vector machine (RSVM) is a typical pairwise method of learning to rank, which is effective in ranking problems. However, the training speed of RSVMs are not satisfactory, especially when solving large-scale data ranking problems. Recent years, many-core processing units (graphics processing unit (GPU), Many Integrated Core (MIC)) and multi-core processing units have exhibited huge superiority in the parallel computing domain. With the support of hardware, parallel programming develops rapidly. Open Computing Language (OpenCL) and Open Multi-Processing (OpenMP) are two of popular parallel programming interfaces. The authors present two high-performance parallel implementations of RSVM, an OpenCL version implemented on multi-core and many-core platforms, and an OpenMP version implemented on multi-core platform. The experimental results show that the OpenCL version parallel RSVM achieved considerable speedup on Intel MIC 7110P, NVIDIA Tesla K20M and Intel Xeon E5-2692v2, and it also shows good portability.