Major Depressive Disorder (MDD) is a common mental disorder, and making an early and accurate diagnosis is crucial for effective treatment. Functional Connectivity Network (FCN) constructed based on functional Magnetic Resonance Imaging (fMRI) have demonstrated the potential to reveal the mechanisms underlying brain abnormalities. Deep learning has been widely employed to extract features from FCN, but existing methods typically operate directly on the network, failing to fully exploit their deep information. Although graph coarsening techniques offer certain advantages in extracting the brain's complex structure, they may also result in the loss of critical information. To address this issue, we propose the Multi-Granularity Brain Networks Fusion (MGBNF) framework. MGBNF models brain networks through multi-granularity analysis and constructs combinatorial modules to enhance feature extraction. Finally, the Constrained Attention Pooling (CAP) mechanism is employed to achieve the effective integration of multi-channel features. In the feature extraction stage, the parameter sharing mechanism is introduced and applied to multiple channels to capture similar connectivity patterns between different channels while reducing the number of parameters. We validate the effectiveness of the MGBNF model on multiple classification tasks and various brain atlases. The results demonstrate that MGBNF outperforms baseline models in terms of classification performance. Ablation experiments further validate its effectiveness. In addition, we conducted a thorough analysis of the variability of different subtypes of MDD by multiple classification tasks, and the results support further clinical applications.

Major depressive disorder (MDD) is a common neuropsychiatric disorder, yet its underlying physiological mechanisms remain unclear, limiting diagnostic advances. Functional connectivity (FC) derived from resting-state functional magnetic resonance imaging (rs-fMRI), when combined with deep learning methods, has shown promise as diagnostic biomarker. Currently, most FC-based diagnostic methods rely on graph structures modeled by FC, and are limited to pairwise interactions between brain regions. Hypergraph representations enable the characterization of higher-order interactions across multiple regions. However, existing hypergraph models ignore the directionality of these interactions, thus limiting their ability to capture complex neural dynamics. To address these limitations, this study proposes a dynamic weight aware directed hypergraph learning method - dwDHGL, for MDD identification and subtype analysis. dwDHGL captures asymmetric causal interactions by modeling temporal lag effects and constructs a directed hypergraph network(DHN). It further utilizes a self-attention mechanism to dynamically learn inter node interactions during message passing and adaptively differentiate node importance. A node weight aware directed hypergraph convolution is designed to aggregate features based on hyperedge directions, incorporating dynamic weights to enhance representation learning. The proposed method is evaluated on the large-scale REST-meta-MDD dataset, achieving an MDD identification accuracy of 73.75 %, and outperforming existing advanced methods in subtype identification. Furthermore, dwDHGL identifies discriminative directed hyperedges, with the inferior frontal gyrus triangular part emerging as key biomarkers, providing new insights into the neural mechanisms of MDD.

Developing generalizable seizure prediction models across patients is clinically imperative but challenged by significant inter-subject EEG variability. Prevailing methods still yield poor cross-patient accuracy, failing to overcome this barrier. To address this, we propose dual attention learning on crossfrequency phase-amplitude coupling networks (CFPAC-DAL) for generalized seizure prediction. Firstly, cross-frequency PAC brain networks were constructed to quantify neurodynamic couplings; next, a dual-attention mechanism for spatial extraction: intra-graph attention learns node interactions within individual networks, and inter-graph attention captures cross-frequency dependencies between distinct PAC graphs; Finally, temporal dynamics are modeled to capture long-range dependencies in EEG sequences. Evaluations on CHB-MIT and Siena datasets show 98.27%/98.24% patient-specific accuracy and 86.28%/81.15% cross-patient accuracy, outperforming the existing state-of-the-art methods, substantiating that dual-attention on cross-frequency PAC networks encodes neural signatures resilient to individual differences, establishing a new paradigm for clinical-ready prediction.

Log-based anomaly detection is a essential task for ensuring the reliability and performance of software systems. However, the performance of existing anomaly detection methods heavily relies on labeling, while labeling a large volume of logs is highly challenging. To address this issue, many approaches based on transfer learning and active learning have been proposed. Nevertheless, their effectiveness is hindered by issues such as the gap between source and target system data distributions and cold-start problems. In this paper, we propose LogAction, a novel log-based anomaly detection model based on active domain adaptation. LogAction integrates transfer learning and active learning techniques. On one hand, it uses labeled data from a mature system to train a base model, mitigating the cold-start issue in active learning. On the other hand, LogAction utilize free energy-based sampling and uncertainty-based sampling to select logs located at the distribution boundaries for manual labeling, thus addresses the data distribution gap in transfer learning with minimal human labeling efforts. Experimental results on six different combinations of datasets demonstrate that LogAction achieves an average 93.01% F1 score with only 2% of manual labels, outperforming some state-of-the-art methods by 26.28%. Website: https://logaction.github.io

Fusing a sequence of static images captured at various exposures, has shown great potential to approach High Dynamic Range (HDR) imaging. However, in the presence of large motion, mis-alignment is almost inevitable and leads to the ghost artifacts. Besides, factors such as noise in the dark region or color saturation in the over-bright region may also fail to fill local image details to the HDR image. This paper provides a motion-robust multi-exposure fusion model based on Swin Transformer. Particularly, we design feature selection gates, which are integrated with the feature warping layers to detect and block mis-alignments. To reconstruct the missing local details, we exploit the long distance contextual dependency in the exposure-space pyramid by the self-attention mechanism. Extensive numerical and visual evaluation has been conducted on a variety of benchmark datasets. The experiments show that our model achieves superior accuracy and efficacy to existing methods.

Existing demosaicking neural models typically follow the same design principles as those adopted by the rest image restoration tasks; yet one unique & often overlooked problem with demosaicking is it also suffers from the moiré artifact. This observation inspires us to examine the key reason that moiré happens only to demosaicking instead of other image restoration tasks. In this process, We identify the spectral inconsistency concealed in the input of demosaicking neural networks; our findings also clarify the failure of the traditional dense spatiospectral feature aggregation in mitigating spectral inconsistency. Based on the analysis, a new solution is proposed to address moiré while preserving fine image details. In particular, we decouple the traditionally used spatio-spectral feature aggregation into comprehensive spectral aggregation and local spatial aggregation. Throughout a diverse range of experiments with quantitative and qualitative results, our approach is shown capable of significantly reducing artifacts such as moiré and over-smoothness, as well as drastically boosting state-of-the-art performance without increasing computational cost.

Identifying and blocking malicious domains is one of the most direct and effective ways to combat malicious websites. In recent years, tools and technologies such as domain registration services, code repositories, and Domain-Flux have been exploited to generate malicious websites, enabling them to quickly "reincarnate" after being identified and shut down. Detecting methods for such malicious website domains are gradually shifting from the Belief Propagation (BP) algorithm to the Graph Neural Network (GNN) algorithm because the latter excels at aggregating neighboring node information to integrate node representations and graph embeddings. However, cloud services, Content Delivery Network (CDN), and Carrier-Grade Network Address Translation (CGNAT) often lead to benign-malicious co-location for domains, causing miscommunication in messages passing based on domain association relationships, thus reducing detection accuracy. To overcome this limitation, this paper proposes a malicious website domain detection algorithm based on one Multi-Message Passing Heterogeneous Neural Network (HMPN) framework. We apply this method to real-world network traffic scenarios and find that it achieves a Macro-Average F1 score of 91.70%, outperforming the baselines. We further conduct a measurement and analysis of the malicious websites’ industrial characteristics to better explore and detect malicious websites.

Existing demosaicking neural models generally reconstruct the red, green and blue channels by one unified network. This has the advantage of allowing the RGB channels to share latent features. However, it is unnoticed that the training samples are severely unbalanced among the three channels. As a consequence, a unified model trained to fit the red or blue sample distributions is prone to under-fit the green samples. In this paper, we demonstrate the existence of such conflicts, and analyze the reason behind this phenomenon. To solve this disadvantage, we decouple the traditional three-in-one demosaicking to three independent but collaborative sub-tasks, respecting the distribution of samples in each spectrum. We show that, by our decoupling strategy, a very simple architecture can achieve high accuracy with high efficiency. For substantiation, we construct a model that has merely 1.3M parameters in total, and compares its accuracy to state of the art methods of similar or larger model size, across a wide variety of benchmark datasets. The proposed independent and collaborative architecture gains performance improvement over comparable models in both accuracy and inference time.

To accomplish computationally intensive visual tasks on mobile devices with limited memory and computation capability, a common solution is to offload the tasks to the edge with more powerful computing resources. Nevertheless, due to the extra communication, even edge computing is challenged by the growing demands for real-time interactive tasks (e.g., VR and AR), which are generally performed by deep CNNs with heavy computation. In this paper, we design a novel system called HRCache to reduce the end-to-end latency by effectively recompressing video data and reusing cached inference results on the edge. As similar image regions have been indicated in the offloading video coding, the edge can approximately infer these regions with results stored on the edge to save the time for calculating these regions. Moreover, since the data in these regions do not require calculation, the transmitted data can be further simplified to reduce transmission latency. Furthermore, HRCache can quickly and continuously adjust the coding parameters to adapt the accuracy loss and overall latency with application requirements. Compared with the original offloading schemes, HRCache significantly reduces the average latency, about 13.60% to 18.83%, at little accuracy loss of 1.25% in Top-1 accuracy for classification and 0.135 in IoU for object detection.

The research on semantic similarity of short text plays an important role in machine translation, emotion analysis, information retrieval and other AI business applications. However, according to existing short text similarity research, the characteristics of ambiguous vocabularies are difficult to be effectively analyzed, the solution of the problem caused by words order needs to be further optimized as well. This paper proposes a short text semantic similarity calculation method that combines BERT and time warping distance algorithm, in order to solve the problem of vocabulary ambiguity. The model first uses the pre trained Bert model to extract the semantic features of the short text from the whole level, and obtains a 768 dimensional short text feature vector. Then, it transforms the extracted feature vector into a point sequence in space, uses the CTW algorithm to calculate the time warping distance between the curves connected by the point sequence, and finally uses the weight function designed by the analysis, according to the smaller the time warpage distance is, the higher the degree of small similarity is, to calculate the similarity between short texts. The experimental results show that this model can mine the feature information of ambiguous words, and calculate the similarity of short texts with lexical ambiguity effectively. Compared with other models, it can distinguish the semantic features of ambiguous words more accurately.

With the emergence of the generation-based steganography, the traditional text steganalysis methods show the unsatisfactory detection performance as the manually extracted features are simple and non-universal. The recently proposed deep learning-based text steganalysis methods can obtain the great detection accuracy by extracting the high-level features. In this letter, a hybrid text steganalysis method (R-BILSTM-C) is proposed through combining the advantages of Bidirectional Long Short Term Memory Recurrent Neural Network (Bi-LSTM) and Convolutional Neural Network (CNN). The proposed method can efficiently capture both local features and long-term semantic information from text to improve the detection accuracy. In the proposed method, the Bi-LSTM architecture is used to capture the long-term semantic information of texts. And the asymmetric convolution kernels with different sizes are applied to extract the local relationship between words. In addition, the high dimensional semantic feature space is visualized. Experimental results show that the proposed method adapts to the different steganographic algorithms efficiently, and achieves the comparable or superior detection performance for the various sentence lengths compared with other state-of-the-art text steganalysis methods.

Digital cameras that use Color Filter Arrays (CFA) entail a demosaicking procedure to form full RGB images. To digital camera industry, demosaicking speed is as important as demosaicking accuracy, because camera users have been accustomed to viewing captured photos instantly. Moreover, the cost associated with demosaicking should not go beyond the cost saved by using CFA. For this purpose, we revisit the classical Hamilton-Adams (HA) algorithm, which outperforms many sophisticated techniques in both speed and accuracy. Our analysis shows that the HA pipeline is highly efficient to exploit the originally captured data, but its oversimplified inter- and intra-channel smoothness formulation hinders its accuracy. We therefore propose a very low cost edge sensing scheme, which guides demosaicking by a logistic functional of the difference between directional variations. We extensively compare our algorithm with 27 demosaicking algorithms by running their open source codes on benchmark datasets. Compared to methods of similar computational cost, our method achieves substantially higher accuracy; Whereas compared to methods of similar accuracy, our method has significantly lower cost. On test images of currently popular resolution, the quality of our algorithm is comparable to top performers, yet our speed is tens of times faster. Source code for this work will be released with paper publication.

This paper proposes a new image signature, called Compass Rose, which is particularly suited to optical flow computation. It is differentiable, fast to compute, and robust to additive illumination changes, translation, and fast rotation. We design a sparse flow computation system based on the invariance of the Compass Rose signatures. This is then extended to dense motion estimation by the addition of an optional diffusion step. Quantitative testing on several benchmark sequences shows the Compass Rose attains higher accuracy than the traditional flow signatures under a range of conditions. Finally, we demonstrate its application to human motion estimation, which is challenging for optical flow methods due to fast limb rotation.

Pyramidal Lucas-Kanade (LK) optical flow is a real-time registration technique widely employed by a variety of cutting edge consumer applications. Traditionally, the LK algorithm is applied selectively to image feature points that have strong spatial variation, which include outliers in textured areas. To detect and discard the falsely selected features, previous methods generally assess the goodness of each feature after the flow computation is completed. Such a screening process incurs additional cost. This paper provides a handy (but not obvious) tool for the users of the LK algorithm to remove false features without degrading the algorithm's efficiency. We propose a confidence predictor, which evaluates the ill-posedness of an LK system directly from the underlying data, at a cost lower than solving the system. We then incorporate our confidence predictor into the course-to-fine LK registration to dynamically detect false features and terminate their flow computation at an early stage. This improves the registration accuracy by preventing the error propagation and maintains (or increases) the computation speed by saving the runtime on false features. Experimental results on state-of-the-art benchmarks validate that our method is more accurate and efficient than related works.

This paper proposes the use of an adaptive locally oriented coordinate frame when calculating an optical flow field. The coordinate frame is aligned with the least curvature direction in a local window about each pixel. This has advantages to both fitting the flow field to the image data and in imposing smoothness constraints between neighboring pixels. In terms of fitting, robustness is obtained to a wider variety of image motions due to the extra invariance provided by the coordinate frame. Smoothness constraints are naturally propagated along image boundaries which often correspond to motion boundaries. In addition, moving objects can be efficiently segmented in the least curvature direction. We show experimentally the benefits of the method and demonstrate robustness to fast rotational motion, such as what often occurs in human motion.

In the virtualized environment, memory can be efficiently utilized if the dynamic memory demands of virtual machines can be estimated at runtime. An efficient memory estimator should report the appropriate size of the memory which can be made full use of by the virtual machine while keeping reasonable performance. However, the appropriate size is hard to be estimated accurately with low overhead. This paper presents a memory demand estimator based on the guest operating system behaviors architecturally visible to the virtual machine monitor, and it can accurately reports the expected appropriate memory size with negligible overhead. The estimator consists of two components which respectively, track the amount of the memory residing in virtual address space, and the memory used as page cache only accessible in kernel mode. The experimental results show that the estimation error is only 0.4%~2.1%, and the runtime overhead is only 0.8% on average due to no additional memory protection traps are introduced.

Summary form only given. We propose a fast and efficient lossless codec of continuous-tone images, SPEDIC (simple predictor and edge detector based image codec), which is uniquely suitable for coding screen updates generated by multimedia applications in thin client computing systems. A codec for thin client computing should take account of the tradeoff between compression ratio and coding complexity, because screen update images in thin client computing should be sent from server to client in a timely fashion. We argue that by combining similar but simpler building blocks of the state-of-the-art methods, a little inferior compression ratio of JPEG-LS (the standard for lossless image compression today) can be attained with much lower coding complexity.

Thin-client computing is uniquely suitable for mobile environments, where resource-poor devices may need to access critical applications over wireless networks. In thin-client computing, applications run on a powerful server, which sends screen updates to the client in real time. However, fast and efficient coding methods for screen updates of multimedia applications are very challenging in wireless environments, for the limited bandwidth and the real-time constraint. In this paper, we present SPEDIC (Simple Predictor and Edge Detector based Image Codec), a fast and efficient codec of continuous-tone images, which is appropriate for multimedia applications in wireless thin-client computing. SPEDIC adopts predictive coding, edge coding and run coding, depending on the characteristics of image blocks, and trades off between compression ratio and coding complexity. The experimental results show that SPEDIC provides good compression ratio with very low coding complexity, especially for continuous-tone images, and achieves the best end-to-end latency over wireless networks. Compared with JPEG-LS, the standard of lossless image compression, SPEDIC compresses 2.5 times faster and decompresses 3.2 times faster than it does for a series of continuous-tone images, and is only 6.2% inferior to it in compression ratio.

Summary form only given. This paper describes an adaptive error recovery technique for a DWT-based coding and transmission system, which recovers singularity information as well as continuity. The approach provided in this paper not only interpolates from the boundary, but also maintains the sharpness in background instead of smoothing the discontinuity out. The recovery function is modeled by the numerical solution of a special category of elliptic equation. The difference between adjacent data can be adjusted by the given weights. The choice of the weights allows high flexibility in maintaining contrast, since we can give small weights where we have evidence for edges.