Deep feature embedding aims to learn discriminative features or feature embeddings for image samples which can minimize their intra-class distance while maximizing their inter-class distance. Recent ...state-of-the-art methods have been focusing on learning deep neural networks with carefully designed loss functions. In this work, we propose to explore a new approach to deep feature embedding. We learn a graph neural network to characterize and predict the local correlation structure of images in the feature space. Based on this correlation structure, neighboring images collaborate with each other to generate and refine their embedded features based on local linear combination. Graph edges learn a correlation prediction network to predict the correlation scores between neighboring images. Graph nodes learn a feature embedding network to generate the embedded feature for a given image based on a weighted summation of neighboring image features with the correlation scores as weights. Our extensive experimental results under the image retrieval settings demonstrate that our proposed method outperforms the state-of-the-art methods by a large margin, especially for top-1 recalls.
In this study, we develop a new approach, called zero-shot learning to index on semantic trees (LTI-ST), for efficient image indexing and scalable image retrieval. Our method learns to model the ...inherent correlation structure between visual representations using a binary semantic tree from training images which can be effectively transferred to new test images from unknown classes. Based on predicted correlation structure, we construct an efficient indexing scheme for the whole test image set. Unlike existing image index methods, our proposed LTI-ST method has the following two unique characteristics. First, it does not need to analyze the test images in the query database to construct the index structure. Instead, it is directly predicted by a network learnt from the training set. This zero-shot capability is critical for flexible, distributed, and scalable implementation and deployment of the image indexing and retrieval services at large scales. Second, unlike the existing distance-based index methods, our index structure is learnt using the LTI-ST deep neural network with binary encoding and decoding on a hierarchical semantic tree. Our extensive experimental results on benchmark datasets and ablation studies demonstrate that the proposed LTI-ST method outperforms existing index methods by a large margin while providing the above new capabilities which are highly desirable in practice.
We presented a method based on multigraphs to mathematically define a distribution function in time for the generation of data exchange in a special-purpose communication network. This is needed for ...the modeling and design of communication networks (CNs) consisting of integrated telecommunications and computer networks (ITCN). Simulation models require a precise definition of network traffic communication. An additional problem for describing the network traffic in simulation models is the mathematical model of data distribution, according to which the generation and exchange of certain types and quantities of data are realized. The application of multigraphs enabled the time and quantity of the data distribution to be displayed as operational procedures for a special-purpose communication unit. A multigraph was formed for each data-exchange time and allowed its associated adjacency matrix to be defined. Using the matrix estimation method allowed the mathematical definition of the distribution function values. The application of the described method for the use of multigraphs enabled a more accurate mathematical description of real traffic in communication networks.
Considering that networks based on New Radio (NR) technology are oriented to provide services of desired quality (QoS), it becomes questionable how to model and predict targeted QoS values, ...especially if the physical channel is dynamically changing. In order to overcome mobility issues, we aim to support the evaluation of second-order statistics of signal, namely level-crossing rate (LCR) and average fade duration (AFD) that is missing in general channel 5G models. Presenting results from our symbolic encapsulation point 5G (SEP5G) additional tool, we fill this gap and motivate further extensions on current general channel 5G. As a matter of contribution, we clearly propose: (i) anadditional tool for encapsulating different mobile 5G modeling approaches; (ii) extended, wideband, LCR, and AFD evaluation for optimal radio resource allocation modeling; and (iii) lower computational complexity and simulation time regarding analytical expression simulations in related scenario-specific 5G channel models. Using our deterministic channel model for selected scenarios and comparing it with stochastic models, we show steps towards higherlevel finite state Markov chain (FSMC) modeling, where mentioned QoS parameters become more feasible, placing symbolic encapsulation at the center of cross-layer design. Furthermore, we generate values within a specified 5G passband, indicating how it can be used for provisioningoptimal radio resource allocation.
This paper presents a method for shortening the computation time and reducing the number of math operations required in complex calculations for the analysis, simulation, and design of processes and ...systems. The method is suitable for education and engineering applications. The efficacy of the method is illustrated with a case study of a complex wireless communication system. The computer algebra system (CAS) was applied to formulate hypotheses and define the joint probability density function of a certain modulation technique. This innovative method was used to prepare microsimulation-semi-symbolic analyses to fully specify the wireless system. The development of an iteration-based simulation method that provides closed form solutions is presented. Previously, expressions were solved using time-consuming numerical methods. Students can apply this method for performance analysis and to understand data transfer processes. Engineers and researchers may use the method to gain insight into the impact of the parameters necessary to properly transmit and detect information, unlike traditional numerical methods. This research contributes to this field by improving the ability to obtain closed form solutions of the probability density function, outage probability, and considerably improves time efficiency with shortened computation time and reducing the number of calculation operations.
In this paper, the knowledge based design of digital filter for analysis of spectral components is illustrated. The primary electrical values are analyzed observing of the faults which appear by ...earth short circuits. The main point of view is negative sequence component and higher harmonics in distributed electrical networks. Method of symbolic processing is used to design digital filter that realizes negative sequence component. Designing of digital filter is based on active analog filter in the three-phase electrical networks using bilinear transformation. We use methods of computer algebra tools (Mathematica) to simulate processes in network. Also, we automatically derive properties of digital filter, and the knowledge embedded in symbolic expressions was used to simulate an example system. The characteristics of digital filter are given and annotated using different sampling frequency and entire cases of earth short circuits.
•This paper proposes a new end-to-end pedestrian detection method called the super-resolution detection (SRD) network that aims to solve the low-quality and occlusion problems in intelligent video ...surveillance.•To verify the effectiveness of the proposed SRD algorithm, a new low-quality playground (PG) dataset for pedestrian detection is collected that provides dense and occluded pedestrians with light interference and motion blur in the surveillance images.•Compared with the state-of-the-art methods, our proposed SRD method achieves higher accuracy of pedestrian detection based on the PG dataset. In particular, we demonstrate improved results for more difficult detection cases (light interference and occluded), and overall higher localization precision.
Pedestrian detection has emerged as a fundamental technology for autonomous cars, robotics, pedestrian search, and other applications. Although many excellent object detection algorithms can be used for pedestrian detection, it is still a challenging problem due to the complicated real-world scenarios, e.g., the detection of pedestrians in low-quality surveillance videos. In this paper, we aim to study the challenging topic of pedestrian detection in low-quality images. Low-quality images are interpreted as those taken with a low-resolution camera, heavy weather or a blurred scene, making it difficult to distinguish pedestrians from the background. To solve this problem, we first introduce a dataset called playground (PG) for low-quality image detection. Images from PG are shot using two different camera views, and pedestrian images are taken at different periods, including day and night. The dataset contains a total of 5,752 images with 31,041 annotations. The average size of the pedestrian is 87×41 and the image size is 480 × 640, indicating that these images are taken from very long distances. Then, we propose a super-resolution detection (SRD) network to enhance the resolution of low-quality images that can help distinguish pedestrians from the blurred background. Finally, based on these enhanced images, we adopt and improve the Faster R-CNN network to help relocate occluded pedestrians. Experimental results on this new dataset proved the efficiency and effectiveness of our algorithm on low-quality images.
In traditional vector of locally aggregated descriptors (VLAD) method, the final VLAD vector is reshaped by summing up the residuals between each descriptor and its corresponding visual word. The ...norm of the residuals varies significantly, and it can make “visual burst”. This is caused by a fact that the contribution of each descriptor to VLAD vector is not the same. To address this problem, we add a different weight to each residual such that the contribution of each descriptor to the VLAD vector becomes even to a certain degree. Also, traditional VLAD method only uses the local gradient features of images. Thus it has a low discrimination. In this paper, local color features are extracted and used to the VLAD method. Moreover, we fuse deep features and the multiple VLAD vectors based on local gradient and color information. Also, in order to reduce running time and improve retrieval accuracy, PCA and whitening operations are used for VLAD vectors. Our proposed method is evaluated on three benchmark datasets, i.e., Holidays, Ukbench and Oxford5k. Experimental results show that our proposed method achieves good performance.
Increased efforts in cancer genomics research and bioinformatics are producing tremendous amounts of data. These data are diverse in origin, format, and content. As the amount of available sequencing ...data increase, technologies that make them discoverable and usable are critically needed. In response, we have developed a Semantic Web–based Data Browser, a tool allowing users to visually build and execute ontology-driven queries. This approach simplifies access to available data and improves the process of using them in analyses on the Seven Bridges Cancer Genomics Cloud (CGC; www.cancergenomicscloud.org). The Data Browser makes large data sets easily explorable and simplifies the retrieval of specific data of interest. Although initially implemented on top of The Cancer Genome Atlas (TCGA) data set, the Data Browser’s architecture allows for seamless integration of other data sets. By deploying it on the CGC, we have enabled remote researchers to access data and perform collaborative investigations.
Contrastive Bayesian Analysis for Deep Metric Learning Kan, Shichao; He, Zhiquan; Cen, Yigang ...
IEEE transactions on pattern analysis and machine intelligence,
2023-June-1, 2023-Jun, 2023-6-1, 20230601, Letnik:
45, Številka:
6
Journal Article
Recenzirano
Recent methods for deep metric learning have been focusing on designing different contrastive loss functions between positive and negative pairs of samples so that the learned feature embedding is ...able to pull positive samples of the same class closer and push negative samples from different classes away from each other. In this work, we recognize that there is a significant semantic gap between features at the intermediate feature layer and class labels at the final output layer. To bridge this gap, we develop a contrastive Bayesian analysis to characterize and model the posterior probabilities of image labels conditioned by their features similarity in a contrastive learning setting. This contrastive Bayesian analysis leads to a new loss function for deep metric learning. To improve the generalization capability of the proposed method onto new classes, we further extend the contrastive Bayesian loss with a metric variance constraint. Our experimental results and ablation studies demonstrate that the proposed contrastive Bayesian metric learning method significantly improves the performance of deep metric learning in both supervised and pseudo-supervised scenarios, outperforming existing methods by a large margin.