ABSTRACT In the field of digital signal processing (DSP), the execution of vector operations depends on the optimization of Single Instruction Multiple Data (SIMD) technology. However, manual SIMD vectorization method is complex to develop, poorly portable, and costly to maintain. Therefore, we propose a method of SIMD instruction sequence generation based on LLVM. This method builds a hierarchical instruction generation framework, combines the characteristics of the target architecture, and uses LLVM automatic vectorization tool to gradually convert the vectorized intermediate representation into the target architecture instruction sequence containing SIMD instructions. Experiments on FT‐M7002 hardware platform show that, compared with the vectorization method of manually calling SIMD built‐in functions, the average execution performance of the instruction sequence generated by this method can be improved by up to 70%.

In the rapidly evolving field of unmanned aerial vehicles (UAVs), real-time object detection is crucial for enhancing UAV intelligence. However, existing research often prioritizes complex networks to boost performance, neglecting the inherent computational resource constraints of UAVs. This paper presents FLDet, a family of faster and lighter detectors specifically designed for UAVs. By revisiting the architecture of modern lightweight detectors from a top-down perspective, FLDet offers a novel and comprehensive redesign of the head, neck, and backbone components. Firstly, we propose a Scale Sparse Head (SSH) that utilizes only two heads to detect objects of varying sizes, leveraging scale sparse feature pyramids to balance performance and efficiency. This design provides heuristic guidance for detector architecture development, offering a new paradigm for detector development. Secondly, a Partial Interaction Neck (PIN) is introduced to facilitate partial interaction between different feature scales, thereby reducing computational costs while effectively integrating multi-scale information. Thirdly, inspired by the primate visual pathway, a Stage-Wise Heterogeneous Network (SHN) is presented, employing heterogeneous blocks to capture both local details and contextual information. Finally, we develop a training strategy called Decay Data Augmentation (DDA) to enhance the detector’s generalization capability, leveraging diverse representations generated by strong data augmentation techniques. Experimental results on two challenging UAV-view detection benchmarks, VisDrone2019 and UAVDT, demonstrate that FLDet achieves a state-of-the-art balance among accuracy, latency, and parameter efficiency. In real scenarios tests, the fastest variant, FLDet-N, achieves real-time performance exceeding 52 FPS on an NVIDIA Jetson Xavier NX with only 1.2M parameters. The source code is available athttps://github.com/wsy-yjys/FLDet.

Machine vision plays a crucial role in Earth observation. As a fundamental and challenging task in vision systems, image registration faces new challenges due to increasing collaborative and customization applications. The prevalence of more false matches and low-precision matches is particularly evident in complex and changeable scenarios. In this article, we propose a robust image registration method via topology sort and vision consistence. Initial candidate matches are established via the nearest neighbor ratio of image intensity descriptors. A topological sort across the proximity structure around the point pairs is defined to assess the reliability of candidate matched pairs, effectively eliminating more false matches while retaining highly reliable point pairs. To preserve more point pairs, we develop a spatial visual inspection mechanism to further determine the potential matches from the remaining pairs that do not satisfy the previous topological constraint. During vision inspection, the spatial transformation model is simultaneously estimated. Experimental results on public datasets show that the proposed method outperforms state-of-the-art approaches in both matching accuracy and visual effect.

Hyperspectral image classification methods based on graph convolution network have received extensive attention. However, the traditional distance metric is difficult to fully represent the spectral variability and uncertainty in hyperspectral images. In order to alleviate this problem, a multi-scale fuzzy graph convolutional network is constructed for hyperspectral image classification. In detail, the SLIC algorithm is used to perform superpixel segmentation of hyperspectral images. Each superpixel is regarded as a graph node, and a fuzzy measurement mechanism is introduced to measure the similarity between two nodes to describe the uncertainty between pixels in the hyperspectral image, so as to construct a fuzzy graph convolution. Subsequently, the fuzzy graph convolution is extended to multi-scale to capture the rich contextual information within the hyperspectral image. In the training process, the pixel-level features are integrated into the superpixel-level graph update process to establish the connection between the pixel level and the superpixel level. Finally, experimental results on two publicly available hyperspectral image datasets show that the proposed network outperforms other representative peers.

Abstract The forward triage characteristic of telemedicine highlights its importance again in the COVID-19 pandemic. Telemedicine can provide timely emergency response in the case of environmental or biological hazards, and the patient’s medical privacy data generated in this process can also accelerate the establishment of models for preventing and treating infectious diseases. However, the reuse process of telemedicine user privacy data based on federated learning also faces significant challenges. Differences in regions, economic levels, and grades lead to heterogeneous data and resource-constrained environments, seriously damaging the federated learning process. Besides, the weak password authentication of medical terminals and eavesdropping attacks on transmission channels may cause illegal access to terminals and platforms and leakage of sensitive data. This paper proposed a telemedicine data secure-sharing scheme based on heterogeneous federated learning. Specifically, we proposed a heterogeneous federated learning scheme with model alignment to guide telemedicine practice through the reuse of telemedicine data; in addition, we designed an SM9 threshold identity authentication scheme to guarantee that the patient’s medical privacy data is protected from leakage during the federated learning process. We evaluated our scheme using two third-party medical datasets. The evaluation results indicate that this scheme can still assist the federated learning process in resisting data heterogeneity and resource constraints with almost no performance cost.

In online display advertising, guaranteed delivery (GD) ads and real-time bidding (RTB) are two main ways to sell impressions for a publisher. While RTB has gained increasing popularity, there is still a proportion of revenue generated from GD ads [1]. Existing mainstream impression allocation models deal with the two delivery ways separately, failing to achieve optimal allocation for multi-objective under multi-constraints, e.g., maximizing gross merchandise volume pre mille (GPM) and revenue per mille (RPM), thus limiting the overall revenue for both the publisher and advertisers. To solve the above problems, we propose a unified guaranteed impression allocation framework to optimally allocate impressions for both GD ads and RTB ads simultaneously. Specifically, we formulate the optimization problem as a non-convex quadratically constrained quadratic programming (QCQP) problem. Then we design an end-to-end unified impression allocation framework to approximately solve the QCQP problem. Furthermore, experiments on real data from Tencent News show that our design significantly increases the overall revenue of both the publisher and advertisers, while achieving much faster convergence than the current state-of-the-art methods.

Quad-tree based coding unit partition in High-Efficiency Video Coding (HEVC) achieved significant coding efficiency improvements, but also brought increasing computational complexity. Especially, design challenges like data dependence, large area cost, and imbalance of processing time of each coding tree unit (CTU), make it hard to achieve a real-time structure for real-time hardware encoder for all CTU sizes. To solve these problems, we proposed a hardware-friendly fast CU decision scheme with multi-stage algorithms for HEVC hardware encoder, aiming at the most complex modules: IME (Integer Motion Estimation), FME/IP (Fractional Motion Estimation/ Intra Prediction), and MD (Mode Decision). Firstly, in IME stage, a zero block detection method based fast CU and PU decision algorithm was presented. Secondly, we presented an estimated RDO (Rate-Distortion Optimization) based algorithm in the Hadamard domain for the early CU decision further in FME/IP stage. Finally, under the condition of hardware computing time limitation of several CU sizes, we proposed a computation time constraint CU fast decision algorithm for MD stage. Experiments demonstrated that, compared with the original HM13.0 implementation, the proposed scheme achieved about 53.9% encoding time saving with merely 2.3% coding performance degradation. What's more, significant area cost and data dependency have been alleviated, which will be more hardware-friendly for HEVC encoder design.

Hyperspectral anomaly detection is an important task in the remote sensing domain. Recently, researchers have shown great interest in deep learning-based methods because they can learn hierarchical, abstract, and high-level representations. However, the latent features learned from the autoencoder (AE) are not always able to reflect the intrinsic structure of hyperspectral data because the locality property is not considered during the learning process. In order to address this problem, a novel manifold constrained AE network (MC-AEN)-based hyperspectral anomaly detection method is proposed in this article. First, the manifold learning method is employed to learn the embedding manifold. Then, the latent representations are learned by an AE network with the learned embedding manifold constraints to preserve the intrinsic structure of hyperspectral data. Finally, the reconstruction errors are calculated to detect anomalies. The global reconstruction error from MC-AEN and the local reconstruction error from the learned latent representations are combined to fully utilize the learned knowledge for better detection performance. We test our proposed algorithm on three different real data sets. Experimental results on these three data sets show the superiority of our proposed method.