基于改进MobileFaceNet的人脸识别方法北大核心

针对目前大多数人脸识别算法参数多、计算量大,难以部署到移动端和嵌入式设备中的问题,提出了一种基于改进MobileFaceNet的人脸识别方法。通过对MobileFaceNet模型结构的调整,将bottleneck模块优化为sandglass模块,改良深度卷积和逐点卷积的相对位置,适当增大sandglass模块的输出通道数,从而减少特征压缩时的信息丢失,增强人脸空间特征的提取。实验结果表明:改进后的方法在LFW测试数据集上准确率达99.15%,模型大小和计算量分别仅为原算法的61%和45%,验证了所提方法的有效性。     

一种基于FCOS神经网络的小建筑物目标检测方法北大核心

提出了一种基于FCOS神经网络的小建筑物目标检测算法,针对FCOS模型在特征提取阶段提取到的小建筑物目标特征较少问题,引入多尺度检测和可变形卷积方式,加强网络对小建筑物目标的特征提取能力,并通过改进后的SGE注意力机制降低特征图中的干扰噪声权重。改进后的网络可以提取到更多的小建筑物目标特征,对环境干扰噪声的鲁棒性更强。在自己搭建的数据集上进行了实验测试,结果表明,在相同环境下网络改进后建筑物的整体检测准确率提升了1.7%,其中对小建筑物目标提升了3.6%,减少了小建筑物目标漏检、误检的问题。     

Mutual information maximizing GAN inversion for real face with identity preservation

Recent generative adversarial networks (GANs) have yielded remarkable performance in face image synthesis. GAN inversion embeds an image into the latent space of a pretrained generator, enabling it to be used for real face manipulation. However, current inversion approaches for real faces suffer the dilemma of initialization collapse and identity loss. In this paper, we propose a hierarchical GAN inversion for real faces with identity preservation based on mutual information maximization. We first use a facial domain guaranteed initialization to avoid the initialization collapse. Furthermore, we prove that maximizing the mutual information between inverted faces and their identities is equivalent to minimizing the distance between identity features from inverted and original faces. Optimization for real face inversion with identity preservation is implemented on this mutual information-maximizing constraint. Extensive experimental results show that our approach outperforms state-of-the-art solutions for inverting and editing real faces, particularly in terms of face identity preservation.     

Collaborative Distribution Alignment for 2D image-based 3D shape retrieval

Retrieving 3D shapes with 2D images has become a popular research area nowadays, and a great deal of work has been devoted to reducing the discrepancy between 3D shapes and 2D images to improve retrieval performance. However, most approaches ignore the semantic information and decision boundaries of the two domains, and cannot achieve both domain alignment and category alignment in one module. In this paper, a novel Collaborative Distribution Alignment (CDA) model is developed to address the above existing challenges. Specifically, we first adopt a dual-stream CNN, following a similarity guided constraint module, to generate discriminative embeddings for input 2D images and 3D shapes (described as multiple views). Subsequently, we explicitly introduce a joint domain-class alignment module to dynamically learn a class-discriminative and domain-agnostic feature space, which can narrow the distance between 2D image and 3D shape instances of the same underlying category, while pushing apart the instances from different categories. Furthermore, we apply a decision boundary refinement module to avoid generating class-ambiguity embeddings by dynamically adjusting inconsistencies between two discriminators. Extensive experiments and evaluations on two challenging benchmarks, MI3DOR and MI3DOR-2, demonstrate the superiority of the proposed CDA method for 2D image-based 3D shape retrieval task.     

Multi-scale and multi-patch transformer for sandstorm image enhancement

Sandstorm is a meteorological phenomenon common in arid and semi-arid regions. A sandstorm can carry large volumes of sand unexpectedly, which leads to severe color deviations and significantly degraded visibility when an image is taken in such a scenario. However, existing image enhancement methods cannot enhance sandstorm images well due to the challenging degradations and the scarcity of sandstorm training data. In this paper, we propose a Transformer with rotary position embedding to perform sandstorm image enhancement via building multi-scale and multi-patch dependencies. Our key insights in this work are 1) a multi-scale Transformer can globally eliminate the color deviations of sandstorm images via aggregating global information, 2) a multi-patch Transformer can recover local details well via learning the spatial variant degradations, and 3) a U-shape Transformer with rotary position embedding as the core unit of multi-scale and multi-patch Transformer can effectively build the long-range dependencies. We also contribute a real-world Sandstorm Image Enhancement (SIE) dataset including 1,400 sandstorm images with different degrees of degradations and various scenes. Experiments performed on synthetic images and real-world sandstorm images demonstrate that our proposed method not only obtains visually pleasing results but also outperforms state-of-the-art methods qualitatively and quantitatively.     

碳中和背景下配电网差异化节能降耗潜力评估分

为了提高配电网差异化节能降耗效果，解决现有潜力评估方法存在的应用性能差的问题，提出碳中和背景下配电网差异化节能降耗潜力优化评估方法。根据配电网的空间结构，构建相应的等值电路模型。在该模型下，从设备损耗和运行附加损耗2个方面计算配电网的损耗量。根据损耗量计算结果，确定配电网差异化碳中和节能降耗方式。从静态和动态2个角度设置潜力评估指标，通过指标数据处理、指标权重求解等步骤，得出配电网差异化节能降耗潜力的综合量化评估结果。将设计潜力评估方法应用到配电网的差异化节能降耗改造工作中，能够有效降低配电网的实际线损量、降低区域损耗费用，并具有较高的应用价值。     

爆破位置对深凹露天矿山炮烟扩散的影响研究

深凹露天矿山由于其特殊的结构,爆破产生的炮烟扩散稀释较为困难,严重危害生产作业人员的生命安全与健康。基于实际矿山构建了深凹露天矿山的二维物理及数学模型,采用非稳态数值分析方法研究了不同爆破位置下,深凹露天矿山采坑内爆破炮烟的扩散规律。研究结果表明：不同爆破位置下,露天采坑内均出现复环流,爆破点位置是影响露天采坑内风流结构特征的重要因素；露天采坑内的炮烟最高浓度均随着时间变化而逐渐下降,但下降的速率逐步减小,呈现三个阶段的下降趋势；爆破位置位于背风侧时露天采坑内的炮烟最高浓度和降至安全浓度所需时间远高于迎风侧三个爆破位置；随着背风侧爆破点距采坑底部距离的减小,炮烟最高浓度及降至安全浓度所需时间先降低后增加,炮烟最高浓度及降至安全浓度所需时间随着迎风侧爆破位置距采坑底部距离的减小而增加。研究结果对于指导深凹露天矿山企业合理组织爆破后的生产作业和保障作业人员安全具有重要意义。     

一种直升机避障电力线检测方法

电力线是一类形状细长、特征稀疏、随着视角的变化容易混淆在大量背景信息中的特殊障碍物，常规电力线检测识别算法得到的目标框对电力线所在位置的估计不够准确。为此，提出了一种相对角度估计方法，基于常规电力线目标检测与识别算法，并结合电力线相对角度估计，从而提高电力线的检测识别过程中所在位置的精度。相比电力线绝对角度回归的方法，提出的相对角度估计方法容易训练易收敛，计算量小，适用于实时性要求较高的应用场合。     

疫情下电磁场理论课程的网络教学

本文主要总结了新冠疫情期间作者的电磁场理论课程在线教学经验。对比分析了录播和直播的优缺点后,选择录播教学方式。基于超星网络教学平台,展示了录播网络教学的具体措施,包括网上答疑和学习效果检查以及在线批改作业等。给出了网络教学可以为线下教学继续使用的方法和手段,为疫情结束后的正常教学提供了新的网络教学补充措施。     

5G核心网业务模型的智能化预测研究

摘 要：核心网业务模型的建立是5G网络容量规划和网络建设的基础，通过现有方法得到的理论业务模型是静态不可变的且与实际网络存在偏离。为了克服现有5G核心网业务模型与现网模型适配性较差以及规划设备无法满足用户实际业务需求的问题，提出了一种长短期记忆（long short-term memory，LSTM）网络与卷积LSTM （convolution LSTM，ConvLSTM）网络双通道融合的 5G 核心网业务模型预测方法。该方法基于人工智能（artificial intelligence，AI）技术以实现高质量的核心网业务模型的智能预测，形成数据反馈闭环，实现网络自优化调整，助力网络智能化建设。