基于域与样例平衡的多源迁移学习方法

针对如何有效使用多源域的决策知识去预测目标域样例标签的问题，提出一种平衡域与样例信息的多源迁移学习算法.为实现上述目的，本文提出了一种基于域与样例平衡的多源迁移学习方法（Multi-source Transfer Learning by Balancing both Domains and Instances，MTL-BDI）.该方法的基本思想是将域层面和样例层面的双加权平衡项嵌入到迁移学习的原始目标函数中，然后利用交替优化技术对提出的目标函数进行有效求解.在文本和图像数据集上的大量实验表明，该方法在分类精度方面确实优于现有的多源迁移学习方法MCC-SVM（Multiple Convex Combination of SVM）、A-SVM（Adaptive SVM）、Multi-KMM（Multiple Kernel Mean Matching）和DAM（Domain Adaptation Machine）.     

Cross-dataset person re-identification method based on multi-pool fusion and background elimination network

The existing cross-dataset person re-identification methods were generally aimed at reducing the difference of data distribution between two datasets,which ignored the influence of background information on recognition performance.In order to solve this problem,a cross-dataset person re-ID method based on multi-pool fusion and background elimination network was proposed.To describe both global and local features and implement multiple fine-grained representations,a multi-pool fusion network was constructed.To supervise the network to extract useful foreground features,a feature-level supervised background elimination network was constructed.The final network loss function was defined as a multi-task loss,which combined both person classification loss and feature activation loss.Three person re-ID benchmarks were employed to evaluate the proposed method.Using MSMT17 as the training set,the cross-dataset mAP for Market-1501 was 35.53%,which was 9.24% higher than ResNet50.Using MSMT17 as the training set,the cross-dataset mAP for DukeMTMC-reID was 41.45%,which was 10.72% higher than ResNet50.Compared with existing methods,the proposed method shows better cross-dataset person re-ID performance.     

MTNet: Mutual tri-training network for unsupervised domain adaptation on person re-identification

The existing unsupervised domain adaptation (UDA) methods on person re-identification (re-ID) often employ clustering to assign pseudo labels for unlabeled target domain samples. However, it is difficult to give accurate pseudo labels to unlabeled samples in the clustering process. To solve this problem, we propose a novel mutual tri-training network, termed MTNet, for UDA person re-ID. The MTNet method can avoid noisy labels and enhance the complementarity of multiple branches by collaboratively training the three different branch networks. Specifically, the high-confidence pseudo labels are used to update each network branch according to the joint decisions of the other two branches. Moreover, inspired by self-paced learning, we employ a sample filtering scheme to feed unlabeled samples into the network from easy to hard, so as to avoid trapping in the local optimal solution. Extensive experiments show that the proposed method can achieve competitive performance compared with the state-of-the-art person re-ID methods.     

Sample selection for visual domain adaptation via sparse coding

Visual domain adaptation has attracted much attention and has made great achievement in recent years. It deals with the problem of distribution divergence between source and target domains. Current methods mostly focus on transforming images from different domains into a common space to minimize the distribution divergence. However, there are many irrelevant source samples for target domain even after the transformation. In order to eliminate the irrelevant samples, we develop a sample selection algorithm using sparse coding theory. We do the sample selection in a common subspace of source and target data to find as many as relevant source samples. In the common subspace, data characteristics are preserved by using graph regularization. Therefore, we can select the most relevant samples for our target image classification task. Moreover, in order to build a discriminative classifier for the target domain, we use not only the common part of source and target domains learned in the common subspace but also the specific part of target domain. The algorithm can be extended to handle samples from multiple source domains. Experimental results show that our visual domain adaptation method on the image classification tasks can be very effective for the state-of-the-art datasets.     

Dual-path image pair joint discrimination for visible–infrared person re-identification

Because the imaging spectra of infrared images and visible light images are different, there is a huge modal difference between visible light images and infrared ones. Existing methods use image conversion to solve the problem of modal difference between two images, but these methods usually fail to focus on the complete information of images, which lead to the results of cross modal person re-identification are unstable. To solve this problem, we propose a new visible–infrared person re-identification method, called dual-path image pair joint discriminant model (DPJD), which simultaneously optimizes the distance within and between classes, and supervises the network learning to identify feature representations. We generate images with different modalities for the samples, and separately compose the same modality image pair and different modality image pair so as to overcome the inconsistent alignment issues. In addition, we also propose a discriminant module based on dual-path (DMDP) to improve the generation quality and discrimination accuracy of image pairs. Experiments on two benchmark datasets SYSU-MM01 and RegDB demonstrate its effectiveness.     

可迁移测度准则下的协变量偏移修正多源集成方法

迁移学习通过充分利用源域共享知识,实现对目标域的小样本问题求解,然而,对训练和测试样本分布差异测度仍然是该领域的主要挑战。该文针对多源迁移学习算法中,由于源域选择和源域辅助样本选择不当引起的负迁移问题进行研究,提出一种可迁移测度准则下的协变量偏移修正多源集成方法。首先,根据源域和目标域之间的协变量偏移原则,利用联合概率的密度估计,定义辅助样本的可迁移测度,验证目标域和源域在数据空间中标记分布的一致性。其次,在多源域选择阶段,引入非迁移判别过程,提高了源域知识的迁移准确性。最后,在Caltech 256数据集中,验证了Gist特征知识表示和迁移的有效性,分析了多种条件下的辅助样本选择和源域选择的有效性。实验结果表明所提算法可有效降低负迁移现象的发生,获得更好的迁移学习性能     

Joint Feature Representation and Classifier Learning Based Unsupervised Domain Adaption ELM

In the problem of unsupervised domain adaption Extreme learning machine (ELM), the output layer parameters need to have both classification and domain adaptation functions, which often cannot be simultaneously fully utilized. In addition, traditional matching method based on data probability distribution cannot find the common subspace of source and target domains under large difference between domains. In order to alleviate the pressure of double functions of classifier parameters, the entire ELM learning process is mainly divided into two stages: feature representation and adaptive classifier learning, thus a joint feature representation and classifier learning based unsupervised domain adaption ELM model is proposed. In the feature representation stage, the source and target domain data are projected to their respective subspace while minimizing the difference in probability distribution between the two domains. In the adaptive classifier learning stage, the smooth manifold regularization term of target domain is used to improve the parameter adaptive ability. Experiments on six different types of datasets show that the proposed model has higher cross-domain classification accuracy.     

融合注意力机制的雷达欺骗干扰域适应识别方法

针对目前雷达欺骗干扰识别中常规特征识别方法应用受限和训练高性能深度学习模型需要的大量标注样本难以高效获取的问题,该文提出一种基于对抗域适应网络的雷达欺骗干扰识别方法,以改善标签限制;并融合注意力机制残差模块进一步提升识别精度。首先,对雷达接收信号进行时频变换后,应用基于对抗网络思想的域适应技术实现从标注源域样本到未标注目标域样本的迁移识别。其次,通过所设计的空间通道注意力残差模块使网络训练聚焦于时频图全局空间特征和高响应通道,以忽略时频图像中可迁移性低的区域抑制负迁移的产生。在不同源域与目标域雷达欺骗干扰数据集上的实验结果表明了该方法的可行性和有效性。     

基于张量表示的域适配的迁移学习中特征表示方法

本文提出一种新的基于张量表示的域适配迁移学习中的特征表示方法,即融合联合域对齐和适配正则化的基于张量表示的迁移学习特征表示方法.当源域和目标域差异很大时,仅将源域对齐潜在共享空间,会造成数据扭曲过大.为缓解此问题,本文方法提出联合域对齐,即源域和目标域同时对齐共享子空间.并且本文方法将适配正则化引入张量表示空间求解.本文适配正则化包括动态分布对齐和图适配,以缩小域间分布差异和保留样本间流行一致性.最后融合联合域对齐,动态分布对齐和图适配,通过联合优化求解获得共享子空间表示.几个公共的跨域数据集上的大量实验结果表明了本文方法优于其它主流的迁移学习方法,验证了本文方法的有效性.     

Transformer-based global–local feature learning model for occluded person re-identification

Most recent occluded person re-identification (re-ID) methods usually learn global features directly from pedestrian images, or use additional pose estimation and semantic analysis model to learn local features, while ignoring the relationship between global and local features, thus incorrectly retrieving different pedestrians with similar attributes as the same pedestrian. Moreover, learning local features using auxiliary models brings additional computational cost. In this work, we propose a Transformer-based dual-branch feature learning model for occluded person re-ID. Firstly, we propose a global–local feature interaction module to learn the relationship between global and local features, thus enhancing the richness of information in pedestrian features. Secondly, we randomly erase local areas in the input image to simulate the real occlusion situation, thereby improving the model’s adaptability to the occlusion scene. Finally, a spilt group module is introduced to explore the local distinguishing features of pedestrian. Numerous experiments validate the effectiveness of our proposed method.     

Modeling feature distances by orientation driven classifiers for person re-identification

To tackle the re-identification challenges existing methods propose to directly match image features or to learn the transformation of features that undergoes between two cameras. Other methods learn optimal similarity measures. However, the performance of all these methods are strongly dependent from the person pose and orientation. We focus on this aspect and introduce three main contributions to the field: (i) to propose a method to extract multiple frames of the same person with different orientations in order to capture the complete person appearance; (ii) to learn the pairwise feature dissimilarities space (PFDS) formed by the subspaces of similar and different image pair orientations; and (iii) within each subspace, a classifier is trained to capture the multi-modal inter-camera transformation of pairwise image dissimilarities and to discriminate between positive and negative pairs. The experiments show the superior performance of the proposed approach with respect to state-of-the-art methods using two publicly available benchmark datasets.     

基于字典对齐的迁移稀疏编码图像分类

基于稀疏编码的图像分类算法,当源域和目标域间样本服从不同分布时,从源域样本中学习到的字典无法有效对目标域样本进行编码,进而严重影响算法的分类性能。为了解决此问题,提出一种基于字典对齐的迁移稀疏编码(TSC-DA)算法。一方面,通过将字典对齐机制引入稀疏编码模型训练过程中,以减少源域和目标域间样本分布差异;另一方面,采用L2正则化项代替字典约束项,将其转化为无约束优化问题,从而回避了拉格朗日对偶法复杂的求解方式。实验结果表明,TSC-DA能够有效提高目标域的图像分类精度。     

一种基于等距度量学习策略的行人重识别改进算法

为了提高行人重识别距离度量MLAPG算法的鲁棒性,该文提出基于等距度量学习策略的行人重识别Equid-MLAPG算法。 MLAPG算法中正负样本对在映射空间的分布不均衡导致间距超参数受负样本对距离影响更大,因此该文设计的Equid-MLAPG算法要求正样本对映射成为变换空间中的一个点,即正样本对在变换空间中距离为零,使算法收敛时正负样本对距离分布不存在交叉部分。实验表明Equid-MLAPG算法能在常用的行人重识别数据集上取得良好的实验效果,具有更好的识别率和广泛的适用性。     

FuseKin: Weighted image fusion based kinship verification under unconstrained age group

Kinship verification using facial images is mainly performed with a single face sample per person. To perform with a single sample, it is very difficult to specify an age group where kin pairs may have higher similarities. To address the above problem, we propose a novel weighted multi sample fusion (WMSF) method. The proposed WMSF method combines kin signals present in multiple samples per person (MSPP) to form a FuseKin image. To select the most discriminant features from the extracted feature vector, we propose a patch based discriminative analysis (PDA) method. Weights are calculated using the PDA method so as to reduce the discrimination between positive FuseKin pairs. Experiments were conducted on two different datasets which contain multiple face image samples per person, namely Family101 and Family in the Wild (FIW) to validate the performance of the proposed methods. Our method achieves competitive results as compared to other state-of-the-art methods.     

基于元学习的无监督风格迁移算法研究

最近,在生成式对抗网络和足够的非配对训练数据下,无监督领域风格迁移取得了较高的性能。然而，现有的领域迁移框架大多基于庞大的训练数据集，且只能根据训练图像进行特定类别的风格迁移，忽略了其中的学习经验被，使获得的模型不能适应新的领域。文中对传统的非配对循环生成对抗网络Cycle-GAN进行了改进，并使用元学习方法训练了无监督领域的风格迁移问题。另外，文中提出的模型在7个不同的双域迁移任务上证明了其有效性，当对每个新领域进行小样本训练时，该算法均优于传统的风格迁移算法。     

Multi-level mutual supervision for cross-domain Person Re-identification

The challenges of cross-domain person re-identification mainly derive from two aspects: (1) The missing of target data labels. (2) The bias between source domain and target domain. Most of existing works focus on only one problem in the above two or deal with them separately. In this paper, we propose a new approach referred as to multi-level mutual supervision to achieve full utilization of labeled source data and unlabeled target data. Along this approach, we construct a dual-branch framework of which the upper branch is trained with original source data and target data while the lower branch is trained with augmented source data and target data. By applying common-pseudo-label and Maximum Mean Discrepancy (MMD) loss in our framework, the mutual supervision in multi levels is achieved. The results show that our model achieves SOTA performance on multiple popular benchmark datasets.     

Multi-object tracking based on deep associated features for UAV applications

Multi-object tracking (MOT) techniques have been increasingly applied in a diverse range of tasks. Unmanned aerial vehicle (UAV) is one of its typical application scenarios. Due to the scene complexity and the low resolution of moving targets in UAV applications, it is difficult to extract target features and identify them. In order to solve this problem, we propose a new re-identification (re-ID) network to extract association features for tracking in the association stage. Moreover, in order to reduce the complexity of detection model, we perform the lightweight optimization for it. Experimental results show that the proposed re-ID network can effectively reduce the number of identity switches, and surpass current state-of-the-art algorithms. In the meantime, the optimized detector can increase the speed by 27% owing to its lightweight design, which enables it to further meet the requirements of UAV tracking tasks.     

基于共同决策方向矢量的多源迁移及其快速学习方法

多源迁移学习提取了多个相似领域之间有用信息,提高了学习效率,但存在计算核矩阵的空间和时间复杂度较高的问题.提出了一种多源迁移学习方法,该方法基于结构风险最小框架理论,以共同决策方向矢量为基准,将多个相似领域的决策方向矢量嵌入到支持向量机的训练过程中,提高了目标领域分类器的分类性能.并结合核心向量机理论提出了共同决策方向矢量核心向量机,实现对大样本数据集的快速分类学习.模拟和真实数据集实验表明了所提算法的有效性.     

基于空域-分频域混合编码的光学图像加密

为了提高现有光学安全系统的安全性能,首次提出“空域-分频域(分数傅里叶域)”混合编码的光学图像加密方法。该方法对图像分别进行一次菲涅耳变换和分数傅里叶变换,并在图像的菲涅耳域和分数傅里叶域上分别利用随机相位版施行相位编码。在加密过程中,上述两次变换的参数均可以设计成密钥．波长因子也可以引入加密过程,这样系统密钥被设计在两个不同的变换域上,提高了系统的安全性能。计算机仿真试验表明,该加密方法能够成功实现图像的加密和解密,系统对密钥参数十分敏感．非授权认证方在不知道密钥参数的情况下几乎不可能对加密图像正确解密。     

Multi-modality medical image fusion based on separable dictionary learning and Gabor filtering

Sparse representation (SR) has been widely used in image fusion in recent years. However, source image, segmented into vectors, reduces correlation and structural information of texture with conventional SR methods, and extracting texture with the sliding window technology is more likely to cause spatial inconsistency in flat regions of multi-modality medical fusion image. To solve these problems, a novel fusion method that combines separable dictionary optimization with Gabor filter in non-subsampled contourlet transform (NSCT) domain is proposed. Firstly, source images are decomposed into high frequency (HF) and low frequency (LF) components by NSCT. Then the HF components are reconstructed sparsely by separable dictionaries with iterative updating sparse coding and dictionary training. In the process, sparse coefficients and separable dictionaries are updated by orthogonal matching pursuit (OMP) and manifold-based conjugate gradient method, respectively. Meanwhile, the Gabor energy as weighting factor is utilized to guide the LF components fusion, and this further improves the fusion degree of low-significant feature in the flat regions. Finally, the fusion components are transformed to obtain fusion image by inverse NSCT. Experimental results demonstrate the more competitive results of the proposal, leading to the state-of-art performance on both visual quality and objective assessment.