Search

Recent research advances in salient object detection (SOD) could largely be attributed to ever-stronger multi-scale feature representation empowered by the deep learning technologies. The existing SOD deep models extract multi-scale features via the off-the-shelf encoders and combine them smartly via various delicate decoders. However, the kernel sizes in this commonly-used thread are usually “fixed”. In our new experiments, we have observed that kernels of small size are preferable in scenarios containing tiny salient objects. In contrast, large kernel sizes could perform better for images with large salient objects. Inspired by this observation, we advocate the “dynamic” scale routing (as a brand-new idea) in this paper. It will result in a generic plug-in that could directly fit the existing feature backbone. This paper’s key technical innovations are two-fold. First, instead of using the vanilla convolution with fixed kernel sizes for the encoder design, we propose the dynamic pyramid convolution (DPConv), which dynamically selects the best-suited kernel sizes w.r.t. the given input. Second, we provide a self-adaptive bidirectional decoder design to accommodate the DPConv-based encoder best. The most significant highlight is its capability of routing between feature scales and their dynamic collection, making the inference process scale-aware. As a result, this paper continues to enhance the current SOTA performance. Both the code and dataset are publicly available at https://github.com/wuzhenyubuaa/DPNet. [ABSTRACT FROM AUTHOR]

We investigate privacy-preserving, video-based action recognition in deep learning, a problem with growing importance in smart camera applications. A novel adversarial training framework is formulated to learn an anonymization transform for input videos such that the trade-off between target utility task performance and the associated privacy budgets is explicitly optimized on the anonymized videos. Notably, the privacy budget, often defined and measured in task-driven contexts, cannot be reliably indicated using any single model performance because strong protection of privacy should sustain against any malicious model that tries to steal private information. To tackle this problem, we propose two new optimization strategies of model restarting and model ensemble to achieve stronger universal privacy protection against any attacker models. Extensive experiments have been carried out and analyzed. On the other hand, given few public datasets available with both utility and privacy labels, the data-driven (supervised) learning cannot exert its full power on this task. We first discuss an innovative heuristic of cross-dataset training and evaluation, enabling the use of multiple single-task datasets (one with target task labels and the other with privacy labels) in our problem. To further address this dataset challenge, we have constructed a new dataset, termed PA-HMDB51, with both target task labels (action) and selected privacy attributes (skin color, face, gender, nudity, and relationship) annotated on a per-frame basis. This first-of-its-kind video dataset and evaluation protocol can greatly facilitate visual privacy research and open up other opportunities. Our codes, models, and the PA-HMDB51 dataset are available at: https://github.com/VITA-Group/PA-HMDB51 [ABSTRACT FROM AUTHOR]

A new testing method (oscillating wave method) has been proposed as a solution to the problem of short maintenance time in the traction power supply system, which could complete tests of winding and insulation via one connection and improve the maintenance efficiency of the transformer. Through the method being proved highly sensitive to winding faults, the variation under different winding states, however, remains unclear. This article aims to analyze oscillating wave signatures in the occurrence of axial displacement (AD). First, the modeling method will be proposed that can simulate oscillating wave signatures in different winding states. Then, the time-domain feature and the new proposed frequency-domain feature are applied to analyze the fault extent and location of AD fault. The results show that this model accurately simulates oscillating wave signatures. Moreover, the time-domain feature [Euclidean distance (ED)] is closely related to fault extent. The combined use of the time-domain feature (ED) and the frequency-domain feature (W-Energy entropy in Z-Score) is effective in the identification of fault location. These features, therefore, can provide the reference for winding state evaluation. [ABSTRACT FROM AUTHOR]

Repetitive or ambiguous environment, where structures are highly similar and distinct geometric features are not sufficient, is one of the critical and challenging scenarios for mobile robots to perform global localization or simultaneous localization and mapping (SLAM) tasks. The robots are easy to get lost or mismatched to wrong places in such environments. The existing solutions either rely heavily on pre-installed infrastructures which are inflexible and expensive or rely on single sensor-based global localization methods whose initialization module is not capable enough to provide distinctive information. Thus, this article proposes a hierarchical probabilistic framework that addresses the problem of infrastructure-free mobile robot global localization in GPS-challenged repetitive environments by leveraging both the measured magnetic field (MF) and Light Detection and Ranging (LiDAR) information. The proposed hierarchical system mainly consists of: 1) coarse localization: MF-based localization and 2) fine localization: LiDAR-based localization (LBL). The ambient MF functions as a substitute for the GPS signal, which is considered not available or severely challenged in indoor/semi-indoor local environments. Based on the pre-built MF database, multiple coarse candidate robot poses can first be determined by the proposed initial pose estimation algorithm, which incorporates multivariate Gaussian observation model and random sample consensus (RANSAC) algorithm. Then, using the obtained multiple candidate poses as initialization, the robot can be localized more accurately by LiDAR-based fine localization. Extensive real-world experiments demonstrate that the proposed system achieves over 95% localization success rate and takes less than 2.0-m average traveling distance to localize the mobile robot. [ABSTRACT FROM AUTHOR]

Data-driven fault diagnostics of industrial systems suffer from class-imbalanced problems, which is a common challenge for machine learning algorithms as it is difficult to learn the features of the minority class samples. Synthetic oversampling methods are commonly used to tackle these problems by generating minority class samples to balance the majority and minority classes. Two major issues will influence the performance of oversampling methods which are how to choose the most appropriate existing minority seed samples, and how to synthesize new samples from seed samples effectively. However, many existing oversampling methods are not accurate and effective enough to generate new samples when dealing with high-dimensional faulty samples with different imbalanced ratios, since they do not take these two factors into consideration at the same time. This article develops a novel adaptive oversampling technique: expectation maximization (EM)-based local-weighted minority oversampling technique for industrial fault diagnostics. This method uses a local-weighted minority oversampling strategy to identify hard-to-learn informative minority fault samples and an EM-based imputation algorithm to generate fault samples based on the distribution of minority samples. To validate the performance of the developed method, experiments were conducted on two real-world datasets. The results show that the developed method can achieve better performances, in terms of F-measure, Matthews correlation coefficient (MCC), and Mean (average of F-measure and MCC) values, on multiclass imbalanced fault diagnostics in different imbalance ratios than state-of-arts’ baseline sampling techniques. [ABSTRACT FROM AUTHOR]

In this article, a new testing method for the diagnosis of winding faults based on the resonant frequency of the transformer is proposed. The system that consists of the high-voltage dc power and high-voltage insulated gate bipolar transistor (IGBT) could charge the transformer periodically. The response measured from the bushing of the winding is the oscillating wave that can be used for evaluating the status. Carrying out the new test on a 220-kV three-winding, three-phase transformer in the field, the tests of insulation and the winding have been completed in one test. The calculated frequency verifies the feasibility and applicability of the method in theory and the voltage does not affect the resonant frequency in the test. The three-winding, three-phase model transformer was built to simulate the various parameters of the power supply, which proves that the features of the oscillating waves are brought about by the properties of the transformer. Also, the new testing method provides a new opportunity to do parameter identification of axial displacement and short-circuit (SC) fault. The abovementioned research shows that the proposed method can conduct the test of winding and insulation in one test for a short time, which proves a reference in the field. [ABSTRACT FROM AUTHOR]

In multimedia processing area, image/video frame resizing is recognized as one of the key techniques and has received more and more attention from the research community. It is necessary and very challenging for image/video distribution systems to be able to generate higher resolution video to meet high-end users’ requirements. In this paper, we analyze the bottleneck of hybrid DCT-Wiener scheme firstly and propose a w-norm signal reconstruction scheme from partial frequency coefficients for upsampling based on uncertainty principles. In the proposed scheme, we design a w-norm model to reconstruct up-sized image/video frames by solving convex optimization with constraints. The proposed scheme combines Wiener filter with compressive sensing (CS) theory seamlessly to automatically control the accuracy of both low- and high-frequency components. This scheme can efficiently avoid the fake contours and ringing effects which are the main defects of hybrid DCT-Wiener methods (hybrid methods). Experiments have been carried out to demonstrate that the proposed scheme can achieve significant PSNR and SSIM gains over typical CS schemes and original hybrid method. On the other hand, the proposed scheme controls the accuracy of both low- and high-frequency components by calculating the correlation between reconstructed signals and sampling matrixes during reconstruction which is a good solution to reduce fake contours and ringing effect. This scheme can also achieve noticeable visual quality improvement compared with hybrid methods and typical CS methods. [ABSTRACT FROM AUTHOR]

Video adaptation through transcoding can provide both bit-rate reduction and resolution reduction to meet various requirements from display devices to network links. One important issue in video transcoding is the design of rate control algorithm in order to achieve target bit rate by adjusting certain coding parameters. Among them, proper selection of initial quantization parameter (QP) has been shown to induce noticeable impact on the performance of video transcoding scheme. Current approaches in initialQP determination are either too complicated or lacking adequate accuracy. This paper presents an adaptive QP initialization for H.264/AVC transcoding. First, we carefully build the models of R-MSE and QP-PSNR. Then, we introduce an R-QP model and allocate an optimal target buffer to the first frame by considering its temporal importance. The analysis and the R-QP model lead to a novel scheme to determine the initial QP adaptively to achieve more accurate estimation. Experiments have been carried out to demonstrate that substantial gains in objective quality measures can be consistently obtained. Without increasing complexity in transcoding system, the proposed adaptive initial QP scheme outperforms current existing schemes for various video sequences tested in this research. [ABSTRACT FROM PUBLISHER]

Vehicle networks is a promising technique, which could provide ubiquitous wireless access for vehicles and realize intelligent transportation. Motivated by the recent advances on device-free wireless sensing, this paper explores a novel technique, which could estimate the speed of a vehicle by analyzing its influence on surrounding wireless signals from roadside wireless infrastructures, such as WiFi. This technique could enable vehicle networks with new moving speed sensing ability and evolve it into a large-scale sensing network, which could monitor the traffic status seamlessly. To achieve this goal, in this paper, we propose and formulate a model to characterize the relationship between the phase and amplitude measurements and the vehicle speed. Based on the model, we develop an efficient method to detect the vehicle and estimate its speed accordingly using the frequency-domain information involved in the spectrogram. We carry out extensive experiments utilizing commodity WiFi. The experimental results reveal that the proposed method could estimate the speed with an error of less than 2.6 km/h, which meets the United Nations regulation for speedometer. [ABSTRACT FROM AUTHOR]