基于身份感知模型的素描人脸识别方法

针对素描图像和光学图像之间存在较大的模态差异这一问题,提出了一种基于身份感知模型的素描人脸识别方法,实现跨模态图像生成和素描人脸识别。该方法应用新的感知损失来监督图像生成网络,生成更好的跨模态图像,减少模态差异带来的识别精度损失,并通过三元组损失来正则化类内和类间距离,增强识别模型的性能,用联合训练策略提升素描人脸识别能力。在UoM-SGFSv2、e-PRIP等素描人脸数据集上的实验结果表明,该方法识别效果优于其他对比算法。     

复杂交通环境下多层交叉融合多目标检测

针对目前复杂交通环境下还存在多目标检测精度和速度不高等问题,以特征金字塔网络(Feature Pyramid Network, FPN)为基础,提出了一种多层融合多目标检测与识别算法,以提高目标检测精度和网络泛化能力。首先,采用ResNet101的五层架构将空间分辨率上采样2倍构建自上而下的特征图,按照元素相加的方式将上采样图和自下而上的特征图合并,并构建一个融合高层语义信息与低层几何信息的特征层;然后,根据BBox回归存在训练样本不平衡问题,选择Efficient IOU Loss损失函数并结合Focal Loss提出一种改进Focal EIOU Loss;最后,充分考虑复杂交通环境下的实际情况,进行人工标注混合数据集进行训练。该模型在KITTI测试集上的平均检测精度和速度比FPN分别提升了2.4%和5 frame/s,在Cityscale测试集上平均检测精度和速度比FPN提升了1.9%和4 frame/s。     

基于残差神经网络的无人机射频信号识别

针对民用小型无人机对低空领域安全造成威胁及无人机视觉上识别难的问题,提出了一种无人机射频信号识别的方法,使用深度学习技术学习无人机与控制器之间的射频信号特征来识别无人机。首先将射频信号数据集中的无人机信号进行预处理操作,接着使用残差神经网络进行模型训练,最后使用训练好的网络模型对无人机信号进行识别验证。实验结果表明,该方法识别无人机是否存在的准确率达到99.8%,识别无人机型号的准确率达到91.1%,识别无人机运行模式的准确率达到70.3%,且该方法具备较强的鲁棒性和环境抗干扰能力,性能明显优于基准方法。     

“模式识别”课程线上线下混合教学模式探索与实践

在模式识别课程教学中,针对传统线下课程短板,开展线上与线下相结合的混合教学模式教学改革探索。阐述了本课程教学改革的教学环节设计、教学实施方法、考核方式等,以思维导图和综合案例串联庞杂知识点,采用多样化手段,加强对学生创新思维和能力的培养。教学实践证明,该课程教学改革调动了学生主动性,培育了其创新能力与科学精神,丰富了线上线下混合教学改革的理论研究。     

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”。     

Deep ensemble-based hard sample mining for food recognition

Deep neural networks represent a compelling technique to tackle complex real-world problems, but are over-parameterized and often suffer from over- or under-confident estimates. Deep ensembles have shown better parameter estimations and often provide reliable uncertainty estimates that contribute to the robustness of the results. In this work, we propose a new metric to identify samples that are hard to classify. Our metric is defined as coincidence score for deep ensembles which measures the agreement of its individual models. The main hypothesis we rely on is that deep learning algorithms learn the low-loss samples better compared to large-loss samples. In order to compensate for this, we use controlled over-sampling on the identified ”hard” samples using proper data augmentation schemes to enable the models to learn those samples better. We validate the proposed metric using two public food datasets on different backbone architectures and show the improvements compared to the conventional deep neural network training using different performance metrics.     

Joint frame rate adaptation and object recognition model selection for stabilized unmanned aerial vehicle surveillance

We propose an adaptive unmanned aerial vehicle (UAV)-assisted object recognition algorithm for urban surveillance scenarios. For UAV-assisted surveillance, UAVs are equipped with learning-based object recognition models and can collect surveillance image data. However, owing to the limitations of UAVs regarding power and computational resources, adaptive control must be performed accordingly. Therefore, we introduce a self-adaptive control strategy to maximize the time-averaged recognition performance subject to stability through a formulation based on Lyapunov optimization. Results from performance evaluations on real-world data demonstrate that the proposed algorithm achieves the desired performance improvements.     

基于残差网络和注意力机制的步态识别算法

针对现有行人步态数据集样本较少、多特征融合复杂且识别精准度不高的问题,提出了一种基于残差网络和注意力机制的步态识别算法。首先对每种特征制作标签,然后将表示不同特征的标签做拼接,达到识别多特征的目的。通过迁移ResNet18的预训练模型,并在ResNet18网络第一个和最后一个卷积层后面加入注意力机制,能够加速网络的收敛,提升模型的精准率。〖JP3〗在CASIA-B数据集上进行多次实验,结果表明所提出方法能准确识别行人、状态和不同角度三种步态特征,识别的精准率为97.6%。     

Extracting embedded messages using adaptive steganography based on optimal syndrome-trellis decoding paths

Privacy protection is the key to maintaining the Internet of Things (IoT) communication strategy. Steganography is an important way to achieve covert communication that protects user data privacy. Steganalysis technology is the key to checking steganography security, and its ultimate goal is to extract embedded messages. Existing methods cannot extract under known cover images. To this end, this paper proposes a method of extracting embedded messages under known cover images. First, the syndrome-trellis encoding process is analyzed. Second, a decoding path in the syndrome trellis is obtained by using the stego sequence and a certain parity-check matrix, while the embedding process is simulated using the cover sequence and parity-check matrix. Since the decoding path obtained by the stego sequence and the correct parity-check matrix is optimal and has the least distortion, comparing the path consistency can quickly filter the coding parameters to determine the correct matrices, and embedded messages can be extracted correctly. The proposed method does not need to embed all possible messages for the second time, improving coding parameter recognition significantly. The experimental results show that the proposed method can identify syndrome-trellis coding parameters in stego images embedded by adaptive steganography quickly to realize embedded message extraction.     

增量式块主成分分析的焊缝图像特征提取算法北大核心CSCD

针对焊缝图像特征提取的实时性问题,该文提出一种增量式块主成分分析(incremental block principal component analysis,IBlockPCA)算法,用于焊缝特征主成分的提取。该算法先将焊缝表面图像分割成子图像块并对其进行重构,然后利用提出的IBlockPCA算法对局部块图像进行增量式特征提取,并采用KNN算法对提取的特征主成分进行分类识别;最后在焊缝数据集上进行了算法的性能对比。实验结果表明,该算法在收敛率、分类率及复杂度等方面均优于其他主成分分析(principal component analysis,PCA)算法,其分类识别率为97.5%,其平均处理速度可达50 frame/s,能够满足焊缝表面图像的实时性处理需求。