Existing works mainly focus on crowd and ignore the confusion regions which contain extremely similar appearance to crowd in the background, while crowd counting needs to face these two sides at the same time. To address this issue, we propose a novel end-to-end trainable confusion region discriminating and erasing network called CDENet. Specifically, CDENet is composed of two modules of confusion region mining module (CRM) and guided erasing module (GEM). CRM consists of basic density estimation (BDE) network, confusion region aware bridge and confusion region discriminating network. The BDE network first generates a primary density map, and then the confusion region aware bridge excavates the confusion regions by comparing the primary prediction result with the ground-truth density map. Finally, the confusion region discriminating network learns the difference of feature representations in confusion regions and crowds. Furthermore, GEM gives the refined density map by erasing the confusion regions. We evaluate the proposed method on four crowd counting benchmarks, including ShanghaiTech Part_A, ShanghaiTech Part_B, UCF_CC_50, and UCF-QNRF, and our CDENet achieves superior performance compared with the state-of-the-arts.

AbstractMost of the existing RGB‐T salient object detection methods are usually based on dual‐stream encoding single‐stream decoding network architecture. These models always rely on the quality of fusion features, which often focus on modality‐shared features and overlook modality‐specific features, thus failing to fully utilise the rich information contained in multi‐modality data. To this end, a modality separate tri‐stream net (MSTNet), which consists of a tri‐stream encoding (TSE) structure and a tri‐stream decoding (TSD) structure is proposed. The TSE explicitly separates and extracts the modality‐shared and modality‐specific features to improve the utilisation of multi‐modality data. In addition, based on the hybrid‐attention and cross‐attention mechanism, we design an enhanced complementary fusion module (ECF), which fully considers the complementarity between the features to be fused and realises high‐quality feature fusion. Furthermore, in TSD, the quality of uni‐modality features is ensured under the constraint of supervision. Finally, to make full use of the rich multi‐level and multi‐scale decoding features contained in TSD, the authors design the adaptive multi‐scale decoding module and the multi‐stream feature aggregation module to improve the decoding capability. Extensive experiments on three public datasets show that the MSTNet outperforms 14 state‐of‐the‐art methods, demonstrating that this method can extract and utilise the multi‐modality information more adequately and extract more complete and rich features, thus improving the model's performance. The code will be released at https://github.com/JOOOOKII/MSTNet.

Input: { F 0 n |n ∈ [12, 1] } , inputs for FSD layer “0”. Output: {Pi|i ∈ [0, 3]}, predictions for each layer of FSD. 1: The number of outputs in each layer of FSD is num op = [6, 3, 2, 1] 2: for (i,m) in enumerate(num op) : 3: for n = [m : 1] : 4: if n = m : 5: (dF i n, F i+1 n )← AIM(F i 2n, F i 2n−1) ; 6: else : 7: (dF i n, F i+1 n )← AIM(C⊣⊔(dF i n, F i 2n), F i 2n−1) ; 8: end for 9: Pi = sigmoid(dF i n) 10: end for

Recent memory-based methods have made significant progress for semi-supervised video object segmentation, by explicitly modeling the semantic correspondences between the target frame and the historical ones. However, the indiscriminate acceptance of historical frames into the memory bank and the lack of fine-grained extraction for target objects may incur high latency and information redundancy in these approaches. In this paper, we circumvent the challenges by developing a Structural Feature Reconfiguration Network (SFRNet). The proposed SFRNet consists of two core sub-modules, which are Global-temporal Attention Module (GAM) and Local-spatial Attention Module (LAM). In GAM, we exploit self-attention-based encoders to capture the target objects' temporal context from historical frames. The LAM then reconfigures features with the current frame's spatial structural prior, which reinforces the objectness of foreground objects and suppresses the interference from background regions. By doing so, our model reduces the reliance on the large memory bank containing redundant historical frames, while instead effectively segmenting video objects with spatio-temporal context aggregated from a small set of key frames. We conduct extensive experiments with benchmark datasets, and the results demonstrate our method's favorable performance against the state-of-the-art approaches. The code will be available at https://github.com/zy5037/SFRNet.

Existing RGB-D SOD methods mainly rely on a symmetric two-stream CNN-based network to extract RGB and depth channel features separately. However, there are two problems with the symmetric conventional network structure: first, the ability of CNN in learning global contexts is limited; second, the symmetric two-stream structure ignores the inherent differences between modalities. In this paper, we propose a Transformer-based asymmetric network (TANet) to tackle the issues mentioned above. We employ the powerful feature extraction capability of Transformer (PVTv2) to extract global semantic information from RGB data and design a lightweight CNN backbone (LWDepthNet) to extract spatial structure information from depth data without pre-training. The asymmetric hybrid encoder (AHE) effectively reduces the number of parameters in the model while increasing speed without sacrificing performance. Then, we design a cross-modal feature fusion module (CMFFM), which enhances and fuses RGB and depth features with each other. Finally, we add edge prediction as an auxiliary task and propose an edge enhancement module (EEM) to generate sharper contours. Extensive experiments demonstrate that our method achieves superior performance over 14 state-of-the-art RGB-D methods on six public datasets. Our code will be released at https://github.com/lc012463/TANet.

Email has become an important way for people to communicate with each other. Because of its high security and fast sending and receiving characteristics, it can establish a corporate brand image while being portable and managed. key tool. Therefore, many criminals take advantage of the feature of mailboxes to send false information through mailboxes. This kind of false information mostly appears in the form of cross-modality, which is very confusing and can easily cause extremely bad effects. To solve this problem, this paper proposes a multi-modal false information detection method using contrastive learning pre-training and attention mechanism, uses contrastive learning to align features between different modal data, and uses attention mechanism to achieve different modal features. The interaction between them, the model construction is completed through feature fusion, and finally the accurate detection of fake emails is realized, so as to facilitate the next step of traceability and countermeasures. Compared with the current mainstream methods, the model proposed in this paper has achieved better results in the detection of multi-modal false information, with an increase of more than 6% in accuracy and other aspects, and can achieve more accurate identification and detection of cross-modal false emails.

Virtual Private Network(VPN) can provide a concealed transmission channel for communication and protect the privacy of users. However, it also brings hidden dangers to cybersecurity with its wide application. Malicious behavior or harmful information can be transmitted secretly through VPN tunnels to avoid firewall censorship. Therefore, VPN traffic identification is an important part of ensure cybersecurity. Although many efforts have been made for VPN traffic identification, existing methods mainly focus on supervised learning models. In this paper, we propose an one-class classification model called AAE-DSVDD for VPN traffic identification. First, we introduce Adversarial AutoEncoder(AAE) for preliminary modeling of VPN traffic. AAE can match the aggregated posterior distribution of the hidden layer to an arbitrary prior distribution. It associates the samples with a normal distribution in the hidden space. Secondly, We implement representation learning for VPN traffic via Deep Support Vector Data Description(DSVDD). A standardized method is designed to match the output distribution of DSVDD with the aggregated posterior distribution of AAE. It alleviates the hypersphere collapse problem of DSVDD and improves identification performance. Finally, we verify the abilities of the AAE-DSVDD model on the public dataset ISCXVPN. Compared with other one-class models, AAE-DSVDD achieved the best identification ability for VPN traffic identification. It also improves the recognition ability when identifying strange classes that are not included in the training data.

Existing RGB-D salient object detection methods generally rely on the dual-encoder structure for RGB and depth feature extraction. However, we observe that the encoders in such models are often not adequately trained to obtain superior feature representations. We name this problem the “under-training issue”. To this end, we propose a multi-branch decoding network (MBDNet) to suppress this issue. The MBDNet introduces additional decoding branches with supervision to form a multi-branch decoding (MBD) structure, facilitating the training of the encoders and enhancing the feature representation. Specifically, to ensure the effectiveness of the introduced supervision and improve the performance of additional decoding branches, we design an adaptive multi-scale decoding (AMSD) module. We also design a multi-branch feature aggregation (MBFA) module to aggregate the multi-branch features in MBD to further improve the detection accuracy. In addition, we design an enhancement complement fusion (ECF) module to achieve multi-modality feature fusion. Extensive experiments demonstrate that our MBDNet outperforms other state-of-the-art methods and mitigates the “under-training issue”.

Surrogate models have attracted considerable interest as approximation tools that can save considerable computational resources in various applications. In this study, a recursive surrogate model was developed based on a generalized regression neural network and variable-fidelity surrogate method. The proposed model can continuously improve its prediction accuracy using a novel recursive correction method. Ultimately, a model with sufficient predictive accuracy can be obtained. To verify the performance of the proposed model, we conducted a series of comparative experiments using test functions and an engineering problem. The results showed that the proposed model has better predictive accuracy and robustness than the other benchmark models. Additionally, the impacts of the stopping criteria and spread factor on the performance of the proposed model were investigated, and the time cost associated with the modeling process was analyzed. This model presents a novel option for engineering design optimization. The recursive correction method also provides a new approach for other regression models to further improve their prediction accuracy.

Vision transformers have recently shown strong global context modeling capabilities in camouflaged object detection. However, they suffer from two major limitations: less effective locality modeling and insufficient feature aggregation in decoders, which are not conducive to camouflaged object detection that explores subtle cues from indistinguishable backgrounds. To address these issues, in this paper, we propose a novel transformer-based Feature Shrinkage Pyramid Network (FSPNet), which aims to hierarchically decode locality-enhanced neighboring transformer features through progressive shrinking for camouflaged object detection. Specifically, we propose a nonlocal token enhancement module (NL-TEM) that employs the non-local mechanism to interact neighboring tokens and explore graph-based high-order relations within tokens to enhance local representations of transformers. Moreover, we design a feature shrinkage decoder (FSD) with adjacent interaction modules (AIM), which progressively aggregates adjacent transformer features through a layer-bylayer shrinkage pyramid to accumulate imperceptible but effective cues as much as possible for object information decoding. Extensive quantitative and qualitative experiments demonstrate that the proposed model significantly outperforms the existing 24 competitors on three challenging COD benchmark datasets under six widely-used evaluation metrics. Our code is publicly available at https://github.com/ZhouHuang23/FSPNet.

General Unmanned Aerial Vehicles (UAVs) are widely used through the computer vision functions of onboard cameras and embedded systems. However, due to the limited memory and computing power of embedded devices on the UAV platform, it is a very challenging issue to analyze the real-time scene through the object detection method. To deal with these challenges, this paper compares the performance of different Yolo series models on the Pascal VOC dataset, using mAP and FPS as evaluation metrics, and applies the training results to the XTDrone UAV Simulation Platform for testing. We evaluate YOLOv3, YOLOv3-tiny, YOLOv3-SPP3, YOLOv4, and YOLOv4-tiny on the Pascal VOC benchmark dataset; the mAP of YOLOv4 is 87.48%, which is 14.2% higher than that of YOLOv3. FPS reaches 72, and the test time on the test set is 103.86s; the shortest test time on the validation set is Yolov4-tiny, but the mAP only reaches 50.06%, which is not as good as Yolov3-tiny. This paper compares the performance of five models in the Pascal VOC Dataset and simulates them on the XTDrone platform, and finally concludes that Yolov3-Tiny can meet the requirements of real-time, lightweight, and high precision.

A multifidelity surrogate (MFS) model is a data fusion method for the enhanced prediction of less intensively sampled primary variables of interest (i.e., high-fidelity (HF) samples) with the assistance of intensively sampled auxiliary variables (i.e., low-fidelity (LF) samples). In this article, an MFS model based on the gradient-enhanced radial basis function, termed gradient-enhanced multifidelity surrogate based on the radial basis function (GEMFS-RBF), is proposed to establish a mapping relationship between HF and LF samples. To identify the scaling factor and the undetermined coefficients in GEMFS-RBF, an expanded correlation matrix is constructed by considering the correlations between the acquired samples, the correlations between the gradients, and the correlations between the samples and their corresponding gradients. To evaluate the prediction accuracy of the GEMFS-RBF model, it is compared with the co-Kriging model, multifidelity surrogate based on the radial basis function (MFS-RBF) model, and two single-fidelity surrogate models. The influences of key factors (i.e., the correlations between the HF and LF functions, the subordinations between the sample sets) and the effect of the cost ratio on the performance of GEMFS-RBF are also investigated. It is observed that GEMFS-RBF presents a more acceptable accuracy rate and is less sensitive to the aforementioned factors than the other benchmark models in most cases in this article, which illustrates the practicability and robustness of the proposed GEMFS-RBF model.

How to make the appearance and motion information interact effectively to accommodate complex scenarios is a fundamental issue in flow-based zero-shot video object segmentation. In this paper, we propose an Attentive Multi-Modality Collaboration Network (AMC-Net) to utilize appearance and motion information uniformly. Specifically, AMC-Net fuses robust information from multi-modality features and promotes their collaboration in two stages. First, we propose a Multi-Modality Co-Attention Gate (MCG) on the bilateral encoder branches, in which a gate function is used to formulate co-attention scores for balancing the contributions of multi-modality features and suppressing the redundant and misleading information. Then, we propose a Motion Correction Module (MCM) with a visual-motion attention mechanism, which is constructed to emphasize the features of foreground objects by incorporating the spatio-temporal correspondence between appearance and motion cues. Extensive experiments on three public challenging benchmark datasets verify that our proposed network performs favorably against existing state-of-the-art methods via training with fewer data. The code is released at https://github.com/isyangshu/AMC-Net.

Tracking methods based on correlation filters have recently attracted attention for achieving fast tracking. However, their performance is somewhat limited in long-term tracking tasks, especially in an occlusion situation. To address this issue, we propose a novel structured correlation filter, which depends on coupled interactions between a static model and a dynamic model. Specifically, the static model exploits the star graph to capture spatial information and provides an initial estimation. The dynamic model based on Bayesian inference uses the rough location as a reference to estimate the final target state. Then, the dynamic model provides a feedback to the static regarding their updates. Finally, the dynamic model provides a scale adaptivity mechanism, which makes the proposed tracker effectively deal with not only partial occlusion but also scale variation. Both qualitative and quantitative evaluations on challenging image sequences demonstrate that the proposed method performs favorably against the state-of-the-art tracking algorithms.

Effective integration of contextual information is crucial for salient object detection. To achieve this, most existing methods based on 'skip' architecture mainly focus on how to integrate hierarchical features of Convolutional Neural Networks (CNNs). They simply apply concatenation or element-wise operation to incorporate high-level semantic cues and low-level detailed information. However, this can degrade the quality of predictions because cluttered and noisy information can also be passed through. To address this problem, we proposes a global Recurrent Localization Network (RLN) which exploits contextual information by the weighted response map in order to localize salient objects more accurately. Particularly, a recurrent module is employed to progressively refine the inner structure of the CNN over multiple time steps. Moreover, to effectively recover object boundaries, we propose a local Boundary Refinement Network (BRN) to adaptively learn the local contextual information for each spatial position. The learned propagation coefficients can be used to optimally capture relations between each pixel and its neighbors. Experiments on five challenging datasets show that our approach performs favorably against all existing methods in terms of the popular evaluation metrics.