Search

To achieve ubiquitous wireless connectivity, low earth orbit (LEO) satellite networks have drawn much attention. However, effective beam management is challenging due to time-varying cell load, high dynamic network topology, and complex interference situations. In this paper, under inter-satellite handover frequency and satellite-terrestrial/inter-beam interference constraints, we formulate a practical beam management problem, aiming to maximize the long-term service satisfaction of cells. Particularly, Lyapunov framework is leveraged to equivalently transform the primal problem into multiple single epoch optimization problems, where virtual queue stability constraints replace inter-satellite handover frequency constraints. Since each single epoch problem is NP-hard, we further decompose it into three subproblems, including inter-satellite handover decision, beam hopping design and satellite-terrestrial spectrum sharing. First, a proactive inter-satellite handover mechanism is developed to balance handover frequency and satellite loads. Subsequently, a beam hopping design algorithm is presented based on conflict graphs to achieve interference mitigation among beams, and then a flexible satellite-terrestrial spectrum sharing algorithm is designed to satisfy the demands of beam cells and improve spectral efficiency. Simulation results show that our proposal significantly improves service satisfaction compared with baselines, where the average data queue length of beam cells is reduced by over 50% with affordable handover frequency.

Low earth orbit (LEO) satellite communication based on 3GPP standard is seen as a promising solution to rolling out communication services in areas without terrestrial base stations. However, due to the fast movement of satellites and large beam footprint size, the existing 5G timing advance (TA) estimation mechanism cannot be directly applied when global navigation satellite system is unavailable. In this article, an enhanced TA estimation approach is proposed for LEO satellite communication networks. Specifically, a user-side time-frequency pre-compensation method is introduced at first, which leverages frequency offset measurement on synchronization signal blocks broadcasted by satellites in initial cell search phase. For the random access phase, the upper bound of inter-preamble interference incurred by partial-period cross-correlation operations is derived for a preamble format advised by 3GPP, and it is shown that the interference level is closely related to the square of the number of such operations. Inspired by this result, a cyclic prefix free preamble format is further designed, which features extended guard time, differential power allocation and flexible preamble structure. Numerical results show that our proposal can reduce the missed detection rate of preamble within a beam. Particularly, the missed detection rates of preamble under 32, 48, and 64 users are lower than 1% when SNR = -6 dB, which is a significant improvement compared to baselines. In addition, our proposal can limit the TA estimation error of the detected users to the time length of 25 time-domain sampling points when the subcarrier spacing is 30 kHz and operation frequency is 27 GHz., Comment: 17 pages, 14 figures

Low earth orbit (LEO) satellite communication networks have been considered as promising solutions to providing high data rate and seamless coverage, where satellite beam management plays a key role. However, due to the limitation of beam resource, dynamic network topology, beam spectrum reuse, time-varying traffic arrival and service continuity requirement, it is challenging to effectively allocate time-frequency resource of satellite beams to multiple cells. In this paper, aiming at reducing time-averaged beam revisit time and mitigate inter-satellite handover, a beam management problem is formulated for dynamic LEO satellite communication networks, under inter-cell interference and network stability constraints. Particularly, inter-cell interference constraints are further simplified into off-axis angle based constraints, which provide tractable rules for spectrum sharing between two beam cells. To deal with the long-term performance optimization, the primal problem is transformed into a series of single epoch problems by adopting Lyapunov optimization framework. Since the transformed problem is NP-hard, it is further divided into three subproblems, including serving beam allocation, beam service time allocation and serving satellite allocation. With the help of conflict graphs built with off-axis angle based constraints, serving beam allocation and beam service time allocation algorithms are developed to reduce beam revisit time and cell packet queue length. Then, we further develop a satellite-cell service relationship optimization algorithm to better adapt to dynamic network topology. Compared with baselines, numerical results show that our proposal can reduce average beam revisit time by 20.8% and keep strong network stability with similar inter-satellite handover frequency.

An accurate and automated tissue segmentation algorithm for retinal optical coherence tomography (OCT) images is crucial for the diagnosis of glaucoma. However, due to the presence of the optic disc, the anatomical structure of the peripapillary region of the retina is complicated and is challenging for segmentation. To address this issue, we developed a novel graph convolutional network (GCN)-assisted two-stage framework to simultaneously label the nine retinal layers and the optic disc. Specifically, a multi-scale global reasoning module is inserted between the encoder and decoder of a U-shape neural network to exploit anatomical prior knowledge and perform spatial reasoning. We conducted experiments on human peripapillary retinal OCT images. The Dice score of the proposed segmentation network is 0.820$\pm$0.001 and the pixel accuracy is 0.830$\pm$0.002, both of which outperform those from other state-of-the-art techniques.

We present an object detection framework based on PaddlePaddle. We put all the strategies together (multi-scale training, FPN, Cascade, Dcnv2, Non-local, libra loss) based on ResNet200-vd backbone. Our model score on public leaderboard comes to 0.6269 with single scale test. We proposed a new voting method called top-k voting-nms, based on the SoftNMS detection results. The voting method helps us merge all the models' results more easily and achieve 2nd place in the Google AI Open Images Object Detection Track 2019.

Many advances of deep learning techniques originate from the efforts of addressing the image classification task on large-scale datasets. However, the construction of such clean datasets is costly and time-consuming since the Internet is overwhelmed by noisy images with inadequate and inaccurate tags. In this paper, we propose a Ubiquitous Reweighting Network (URNet) that learns an image classification model from large-scale noisy data. By observing the web data, we find that there are five key challenges, i.e., imbalanced class sizes, high intra-classes diversity and inter-class similarity, imprecise instances, insufficient representative instances, and ambiguous class labels. To alleviate these challenges, we assume that every training instance has the potential to contribute positively by alleviating the data bias and noise via reweighting the influence of each instance according to different class sizes, large instance clusters, its confidence, small instance bags and the labels. In this manner, the influence of bias and noise in the web data can be gradually alleviated, leading to the steadily improving performance of URNet. Experimental results in the WebVision 2018 challenge with 16 million noisy training images from 5000 classes show that our approach outperforms state-of-the-art models and ranks the first place in the image classification task.