基于transformer自适应特征向量融合的图像分类

针对目前基于transformer的图像分类模型直接应用在小数据集上性能较差的问题,本文提出了transformer自适应特征向量融合网络,该网络在特征提取器中将不同阶段的特征进行融合,减少特征信息丢失的同时获得更多不同感受野下的信息,同时利用最大池化来去除特征中的冗余信息,从而使提取的特征更具有判别性。此外,为了充分利用图像的各级特征信息来进行分类预测,本文将网络各阶段产生的特征向量进行融合,使融合后的特征向量更具有表征能力,从而减少网络对大数据集的依赖,使网络在小数据集中也能获得很好的性能。实验表明,本文提出的 算法在数据集Mini-ImageNet-100、CIFAR-100和ImageNet-1k上的TOP-1准确率分别达到了74.22%、85.86%和81.4%。在没有增加计算量的情况下,在baseline上分别提高了6.0%、3.0%和0.1%,且参数量减少了18.3%。本文代码开源在“https://github.com/xhutongxue/afvf”。     

基于改进3D UNet的前列腺MR图像分割

针对前列腺磁共振 (magnetic resonance, MR)图像边缘模糊、对比度较低,灰度值分布不均衡而导致分割精度较差的问题,提出了一种结合双路径注意力(dual path attention,DPA) 和多尺度特征聚合(multi-scale feature aggregation,MFA) 模块的改进3D UNet网络模型。首先,对数据集进行重采样和裁剪处理以适应模型输入。然后,在3D UNet网络的编码器各层引入DPA 并添加残差连接,加强特征的 编码能力。同时,在网络解码器中加入MFA模块,以充分利用空间上下文信息,增强语义信息。最后,在公开数据集PROMISE12上进行验证,所提出的模型的Dice系数为89.90%,Hausdorff 距离为9.37 mm。相比较于其他模型,所提出模型的分割结果更优,且参数量和运算量更少。     

基于提取双选紧密特征的RGB-D显著性检测网络

针对现有算法对不同来源特征之间的交互选择关注度欠缺以及对跨模态特征提取不充分的问题,提出了一种基于提取双选紧密特征的RGB-D显著性检测网络。首先,为了筛选出能够同时增强RGB图像显著区域和深度图像显著区域的特征,引入双向选择模块(bi-directional selection module, BSM);为了解决跨模态特征提取不充分,导致算法计算冗余且精度低的问题,引入紧密提取模块(dense extraction module, DEM);最后,通过特征聚合模块(feature aggregation module, FAM)对密集特征进行级联融合,并将循环残差优化模块(recurrent residual refinement aggregation module, RAM)配合深度监督实现粗显著图的持续优化,最终得到精确的显著图。在4个广泛使用的数据集上进行的综合实验表明,本文提出的算法在4个关键指标方面优于7种现有方法。     

Balancing MXene Surface Termination and Interlayer Spacing Enables Superior Microwave Absorption

Surface chemistry and interlayer engineering determines the electrical properties of 2D MXene. However, it remains challenging to regulate the surface and interfacial chemistry of MXene simultaneously. Herein, simultaneous modulation of Ti₃C₂T_x MXene surface termination and layer spacing by alkali treatment are achieved. The electrical and electromagnetic properties of Ti₃C₂T_x are investigated in detail with respect to KOH and ammonia concentration dependence. A high concentration of KOH caused the Ti₃C₂T_x layer spacing to expand to 13.7 Å and the surface O/F ratio to increase to 33.84. Because of its weaker ionization effect, ammonia provides finer tuning compared to the drastic intercalation of KOH with a thorough sweeping of the F-containing groups. Ti₃C₂T_x is enriched with conductive -OH termination after ammonia treatment, which achieves an effective balance with the increased interlayer resistance. Therefore, NH₃H₂O-Ti₃C₂T_x achieves broad-band impedance matching and exhibits an efficient microwave loss of −49.1 dB at a low thickness of 1.7 mm, with an effective frequency bandwidth of 3.9 GHz. The results herein optimize the electrical properties of Ti₃C₂T_x using surface and interfacial chemistry to achieve broad microwave absorption, providing a framework for enhancing the electromagnetic wave loss of intrinsic MXene.     

Blind omnidirectional image quality assessment with representative features and viewport oriented statistical features

With the development of information technologies, various types of streaming images are generated, such as videos, graphics, Virtual Reality (VR)/omnidirectional images (OIs), etc. Among them, the OIs usually have a broader view and a higher resolution, which provides human an immersive visual experience in a head-mounted display. However, the current image quality assessment works cannot achieve good performance without considering representative human visual features and visual viewing characteristics of OIs, which limited OIs’ further development. Motivated by the above problem, this work proposes a blind omnidirectional image quality assessment (BOIQA) model based on representative features and viewport oriented statistical features. Specifically, we apply the local binary pattern operator to encoder the cross-channel color information, and apply the weighted LBP to extract the structural features. Then the local natural scene statistics (NSS) features are extracted by using the viewport sampling to boost the performance. Finally, we apply support vector regression to predict the OIs’ quality score, and experimental results on CVIQD2018 and OIQA2018 Databases prove that the proposed model achieves better performance than state-of-the-art OIQA models.     

Optimized deep learning-based intrusion detection for wireless sensor networks

The technological innovations and wide use of Wireless Sensor Network (WSN) applications need to handle diverse data. These huge data possess network security issues as intrusions that cannot be neglected or ignored. An effective strategy to counteract security issues in WSN can be achieved through the Intrusion Detection System (IDS). IDS ensures network integrity, availability, and confidentiality by detecting different attacks. Regardless of efforts by various researchers, the domain is still open to obtain an IDS with improved detection accuracy with minimum false alarms to detect intrusions. Machine learning models are deployed as IDS, but their potential solutions need to be improved in terms of detection accuracy. The neural network performance depends on feature selection, and hence, it is essential to bring an efficient feature selection model for better performance. An optimized deep learning model has been presented to detect different types of attacks in WSN. Instead of the conventional parameter selection procedure for Convolutional Neural Network (CNN) architecture, a nature-inspired whale optimization algorithm is included to optimize the CNN parameters such as kernel size, feature map count, padding, and pooling type. These optimized features greatly improved the intrusion detection accuracy compared to Deep Neural network (DNN), Random Forest (RF), and Decision Tree (DT) models.     

HSP-MFL: A High-level Semantic Property driven Multi-task Feature Learning Network for unsupervised person Re-ID

Unsupervised person re-identification aims to distinguish different pedestrians from discriminative representations on the basis of unlabeled data. Currently, most unsupervised Re-ID approaches explore visual representations to generate pseudo-labels for model’s training, which may suffer from background noise and semantic loss. To tackle this problem, this paper proposes a High-level Semantic Property driven Multi-task Feature Learning Network (HSP-MFL) to firstly introduce three high-level semantic properties for unsupervised person Re-ID. Technically, we design a novel Multiple Feature Fusion Module (MFFM) to deeply explore the complex correlation among multiple semantic and visual features to capture the discriminative feature cues, as well as a multi-task training scheme to generate robust fusion features. The architecture is quite simple and does not consume extra labeling costs. Extensive experiments on three datasets demonstrate that both high-level semantic properties and multi-task learning are effective in performance improvement, yielding SOTA mAPs for unsupervised person Re-ID.     

A matching game framework for users clustering and resource allocation with wireless power transfer in a CR-NOMA network

This paper investigates the resource allocation in a massively deployed user cognitive radio enabled non-orthogonal multiple access (CR-NOMA) network considering the downlink scenario. The system performance deteriorates with the number of users who are experiencing similar channel characteristics from the base station (BS) in NOMA. To address this challenge, we propose a framework for maximizing the system throughput that is based on one-to-one matching game theory integrated with the machine learning technique. The proposed approach is decomposed to solve users clustering and power allocation subproblems. The selection of optimal cluster heads (CHs) and their associated cluster members is based on Gale-Shapley matching game theoretical model with the application of Hungarian method. The CHs can harvest energy from the BS and transfer their surplus power to the primary user (PU) through wireless power transfer. In return, they are allowed to access the licensed band for secondary transmission. The power allocation to the users intended for power conservation at CHs is formulated as a probabilistic constraint, which is then solved by employing the support vector machine (SVM) algorithm. The simulation results demonstrate the efficacy of our proposed schemes that enable the CHs to transfer the residual power while ensuring maximum system throughput. The effects of different parameters on the performance are also studied.     

小型化L波段1 000 W GaN内匹配功率管研制

L波段功率单管有增大功率的需求,但会面临体积较大的问题。基于0.5μm工艺研发了GaN高电子迁移率晶体管(HEMT)管芯,单芯功率达到300 W。通过负载牵引仿真提取模型的输入、输出最佳阻抗点。用高介电常数薄膜电路设计L-C网络,拉高芯片的输入输出阻抗,并抵消虚部。用微带电路设计两级阻抗变换的宽带功率分配器及合路器电路,进行四胞管芯合成。内置稳定电路、栅极和漏极供电偏置电路,实现高度集成化、小型化,以及50Ω输入输出阻抗匹配。芯片总栅宽4×40 mm,在漏压50 V、脉宽40μs、占空比4%的测试条件下,在0.96 GHz到1.225 GHz的宽带频段内,输出功率为60 dBm到61.2 dBm,效率为57.9%到72%,饱和功率增益大于14 dB。     

一种实时多尺度目标检测识别算法

无人机目标检测与识别任务中,目标随着飞行高度的改变尺寸发生显著变化。常规目标检测模型中,获取的小目标细节信息有限,检测精度较低;而适用于小目标的实时检测模型往往容易丢失大目标的背景信息,降低大目标的检测精度。针对以上多尺度目标检测识别任务难点,提出一种基于改进特征金字塔网络(Feature Pyramid Network, FPN)结构的实时多尺度目标检测识别模型。该模型通过增加特征金字塔层级覆盖更广的目标尺度,获取更为丰富的目标信息;同时,利用跨连接增加不同尺度特征融合的多样性,降低特征传导距离,保留更加完整的尺度特征来提高模型检测识别多尺度目标的性能。通过实验发现,相比于原始网络结构和相同特征层级的四层特征金字塔结构,加入改进特征金字塔结构的多尺度目标检测模型识别性能得到了提升。