In this ubiquitous age of mobile phones, Android malware is a growing threat to mobile users. The most common malware detection technique is analyzing different features that any Android app encompasses. However, malware developers have become increasingly sophisticated; hence, these Android app features can’t be revealed easily. One way to address this problem is to apply reverse engineering on the Android app executable file, which will eventually decompose the app into low-level source codes. After that, these source codes can be analyzed for subsequent identification of features, which will ultimately help to segregate malware from the benignware. In this article, a step-by-step procedure has been demonstrated to extract four important features (declared permission, intent, component, and hardware/software) from the Android app using reverse engineering. This will open a path for new researchers in this domain.
RNA is a molecule composed of ribonucleotides and plays a crucial role in biological activities. The computational prediction of RNA secondary structures has been a long-standing issue in computational biology. Traditional methods for this problem are based on free energy minimization, but the performance of these methods has reached an upper limit. In recent years, various deep learning-based methods have been proposed, but these models are still primitive and prone to overfitting, resulting in poor performance across RNA families. In this paper, we propose two methods, AttnUFold and TransUFold, which utilize the attention mechanism to enhance the model’s learning ability for the global features of RNA sequences. Additionally, we modify the loss function to cope with sample distribution imbalances and attempt to introduce relevant constraints for RNA folding. Compared with the baseline, the two models have brought improvements in both within- and cross-family tasks. AttnUFold achieved a high F1 score of 0.852 on the ArchiveII dataset, surpassing all traditional and most deep learning methods.
Benefiting from precise perception, real-time prediction and reliable planning, autonomous driving systems have exhibited exceptional performance in research. However, the high complexity and opacity prevent its application in practice. To introduce a user-friendly autonomous driving system, we propose a driving captioner to generate real time description and explanation of self-driving systems in natural language. Specifically, we unify the end-to-end autonomous driving and video captioning tasks into a single yet effective framework by introducing an additional captioning head to describe the action of the vehicle and explain the reasons. Besides, we exploit an effective accelerating method to accelerate the inference process, which decreases the average inference time from 0.670 s to 0.298 s. Through extensive experiments on both simulation datasets and real-world datasets, we show the superior generalization ability and robustness of the proposed framework.
So far, the number of individuals with Tic disorder worldwide has reached 59 million, and the prevalence of the disorder is rapidly increasing globally. In this work, we focus on weakly supervised learning methods for recognizing childhood tic disorders. In situations with limited data availability, we design a relative probability metric based on the characteristics of the data and a multi-phase learning algorithm is proposed based on relative probability in order to efficiently utilize coarse-labeled data in a “from easy to difficult” manner. Furthermore, the effectiveness of our method is validated through ablation experiments. Through extensive experiments on the test dataset, we demonstrate that our method behaves extraordinarily compared to baseline approaches, improving AUC by 3.0%, and facilitating expedited diagnostic assessment for medical practitioners.
Neural network pruning is a widely used approach for reducing the inference cost of deep models in order to deploy on resource-limited settings. However, current pruning works lack attention to information migration from pruned to the remaining part of the Deep Neural Network, and on balancing model performance and compression rate. On these two issues, in this paper, we propose a novel E xplicit I nformation M igration network P runing (EIMP) algorithm. Specifically (1) the constrained gradient update method transfers valid information from redundant networks to the preserved, and (2) the newly designed λ -decay regularization method learns the trade-off between the performance and penalty item. Experiments show that our EIMP algorithm achieves state-of-the-art performance on several datasets with various benchmark network architectures. Notably, EIMP achieves +1.54% better than SOTA on ImageNet.
Fraudulent account detection is essential for businesses and online Internet enterprises, which can help to avoid financial loss and improve user experience. However, conventional solutions suffer from two main challenges which remain unresolved; first, it’s hard to monitor and detect fraud behaviors in real-time, and second, the features of the cheaters keep changing dynamically, which makes it hard to capture the most relevant features for the detection models. In this demonstration, we present a fraudulent account identification system called FAI, which can help to address the above challenges by exploring a multi-granularity sliding window strategy to construct the dynamic features, and both dynamic and static features are embedded together as the input of pre-training models. FAI also provides an interface that allows users to select sets of features in the spatio-temporal dimension flexibly, visualize the feature aggregation results, and assess the quality of fraud detection results. Demo video click here .
GuessWhich is an engaging visual dialogue game that involves interaction between a Questioner Bot (QBot) and an Answer Bot (ABot) in the context of image-guessing. In this game, QBot’s objective is to locate a concealed image solely through a series of visually related questions posed to ABot. However, effectively modeling visually related reasoning in QBot’s decision-making process poses a significant challenge. Current approaches either lack visual information or rely on a single real image sampled at each round as decoding context, both of which are inadequate for visual reasoning. To address this limitation, we propose a novel approach that focuses on visually related reasoning through the use of a mental model of the undisclosed image. Within this framework, QBot learns to represent mental imagery, enabling robust visual reasoning by tracking the dialogue state. The dialogue state comprises a collection of representations of mental imagery, as well as representations of the entities involved in the conversation. At each round, QBot engages in visually related reasoning using the dialogue state to construct an internal representation, generate relevant questions, and update both the dialogue state and internal representation upon receiving an answer. Our experimental results on the VisDial datasets (v0.5, 0.9, and 1.0) demonstrate the effectiveness of our proposed model, as it achieves new state-of-the-art performance across all metrics and datasets, surpassing previous state-of-the-art models.
Visual question answer (VQA) grounding, aimed at locating the visual evidence associated with the answers while answering questions, has attracted increasing research interest. To locate the evidence, most existing methods extract attention maps in an unsupervised manner from pretrained VQA models. As only the text-related objective is considered during training, the attention map coarsely depicts the grounding region, resulting in poor interpretability. A straightforward solution for improving grounding accuracy is leveraging pixel-wise masks as strong supervision. However, precise per-pixel annotation is time-consuming and labor-intensive. To address above issues, this paper presents the weakly-supervised grounding for VQA, which learns an end-to-end Dual Visual-Linguistic Interaction (DaVi) network in a unified architecture with various low-cost annotations, such as click-, scribble- and box-level grounding labels. Specifically, to enable the visual mask prediction, DaVi proposes a language-based visual decoder that extends the previous VQA network. Since the visual decoder is guided with weak labels, we also present a Pseudo Grounding Refinement Module (PGRM) to refine the relatively coarse predictions as an additional constraint. Extensive experiments demonstrate that our weakly supervised DaVi significantly improves grounding performance even under the click-level supervision with one pixel annotation. Scribble-level supervision achieves 92% performance at a dramatically reduced annotation cost compared to its fully supervised counterpart. More essentially, weak visual grounding usually boosts the accuracy of text answers despite using inaccurate supervision.
Although existing image-based methods for 3D human mesh reconstruction have achieved remarkable accuracy, effectively capturing smooth human motion from monocular video remains a significant challenge. Recently, video-based methods for human mesh reconstruction tend to build more complex networks to capture temporal information of human motion, resulting in a large number of parameters and limiting their practical applications. To address this issue, we propose an Efficient Graph Transformer network to Reconstruct 3D human mesh from monocular video, named EGTR. Specifically, we present a temporal redundancy removal module that uses 1D convolution to eliminate redundant information among video frames and a spatial-temporal fusion module that combines Modulated GCN with transformer framework to capture human motion. Our method achieves better accuracy than the state-of-the-art video-based method TCMR on 3DPW, Human3.6M and MPI-INF-3DHP datasets while only using 8.7% of the parameters, indicating the effectiveness of our method for practical applications.
UAVs based on PID controllers are having increasing difficulties in handling complex tasks. Whereas, reinforcement learning-based high-dimensional models provide an important entry point for flight control to handle complex and high-dimensional tasks. In this paper, a neural network controller training framework for outer-loop control is proposed, which is used as a base platform for velocity controller training. Also, a reinforcement learning-based quadrotor neural network speed controller is proposed which maps the state of the UAV to the throttle commands of the rotor for stable control of speed. In addition, this paper employs the idea of curriculum learning to help the UAV adapt to a larger speed tracking range and improve its overall performance. We demonstrate the performance of the trained neural network controller by comparing it with a conventional PI controller in simulations, achieving improvements in both steady-state response time and tracking performance.
As a task aiming to recognize text from images, text recognition is of great significance in both industry and academia. The vast majority of existing text recognition methods use text images with the same styles as training and testing samples. However, when these models encounter images with new styles, their recognition accuracy will be significantly reduced. In this paper, we mainly explore Domain-Generalizable Text Recognition (DGTR), a challenging but meaningful setting focusing on enhancing the generalization ability of text recognition models. For this reason, we propose a practical framework called Sequential Style Consistency Learning (SSC), disentangling the style-specific and task-specific representation. Specifically, our SSC first constructs samples of augmented visual feature sequences, then disentangles the original and augmented feature sequences into style-specific features and task-specific features. To better separate the task-specific representation from the style-specific representation, the Style-Consistency Learning (SCL) is designed for learning the style consistency between original and augmented sequences. The disentangled module and style-consistency learning could provide complementary information for each other. Besides, our SSC is encouraged to meta-learn the style-specific and task-specific features during training based on text images with seen styles, generalizing better to text images with other styles. Numerous experiments and analyses conducted on the benchmark dataset MSDA have shown that SSC can achieve very competitive experimental results compared to state-of-the-art methods.
Lithology identification of rock is one of the main bases for stratigraphic division in geology and plays a very important role in oil and gas exploration. In recent years, with the increasing amount of data obtained by MWD and other methods, it is possible to use artificial intelligence method to dynamically identify lithology based on these data. This paper establishes a formation lithology prediction model based on CNN-LSTM-Attention, predicts formation lithology through drilling parameters and logging data, and verifies the drilling data of a block in Huizhou, South China Sea. Three artificial intelligence methods, convolutional neural network - Long short-term memory neural network -Attention mechanism (CNN-LSTM-Attention), convolutional neural network - long short-term memory neural network (CNN-LSTM) and long short-term memory neural network (LSTM), are compared and analyzed. The results show that the lithology prediction model proposed in this paper has good accuracy and low error, and has certain reliability and practicability.
Pedestrian detection serves as the cornerstone of pedestrian tracking and re-identification, playing a pivotal role in the realm of intelligent transportation. Accurate identification of pedestrians with diverse identities, such as passengers, crew members, and cleaning staff, is of utmost importance in high-security-demand scenarios like airport boarding bridges. The varied poses of pedestrians, occlusions, and small appearance differences pose significant challenges for accurately detecting individuals with different identities in boarding bridge scenarios. Existing object detectors exhibit limited prowess in extracting discriminative features tailored specifically for pedestrians, hampering their ability to fulfill the requirements of precise localization and classification. In this paper, we propose a method based on spatial attention and joint crowd density estimation. By incorporating spatial attention, our network selectively focuses on salient regions corresponding to different pedestrian categories, thereby enhancing classification accuracy. Moreover, through introducing an auxiliary task of crowd density estimation, the supervision of pedestrian head position information is added to the network. This significantly alleviates the missed detection problems caused by perspective distortion and occlusion, leading to significant improvements in detection accuracy. In our study, we use YOLO as the baseline model. The improved model shows a 5.81% increase in mAP and significantly outperforms several common object detectors.
The explosively generated micro-videos on content sharing platforms call for recommender systems to permit personalized micro-video discovery with ease. Recent advances in micro-video recommendation have achieved remarkable performance in mining users’ current preference based on historical behaviors. However, most of them neglect the dynamic and time-evolving nature of users’ preference, and the prediction on future micro-videos with historically mined preference may deteriorate the effectiveness of recommender systems. In this paper, we devise the DMR framework, which comprises: 1) the implicit user network module which identifies sequence fragments from other users with similar interests and extracts the sequence fragments that are chronologically behind the identified fragments; 2) the multi-trend routing module which assigns each extracted sequence fragment into a trend group and update the corresponding trend vector; 3) the history-future trend prediction module jointly uses the history preference vectors and future trend vectors to yield the final click-through-rate. We validate the effectiveness of DMR over multiple state-of-the-art micro-video recommenders on two publicly available real-world datasets. Relatively extensive analysis further demonstrate the superiority of modeling dynamic multi-trend for micro-video recommendation.
Thanks to the recent development of inverse rendering, photorealistic re-synthesis of indoor scenes have brought augmented reality closer to reality. All-angle environment illumination map estimation of arbitrary locations, as a fundamental task in this domain, is still challenging to deploy due to the requirement of expensive depth input. As such, we revisit the appealing setting of illumination estimation from a single image, using a cascaded formulation. The first stage predicts faithful depth maps from a single RGB image using a distortion-aware architecture. The second stage applies point cloud convolution operators that are equivariant to SO(3) transformations. These two technical ingredients collaborate closely with each other, because equivariant convolution would be meaningless without distortion-aware depth estimation. Using the public Matterport3D dataset, we demonstrate the effectiveness of our illumination estimation method both quantitatively and qualitatively. Code is available at https://github.com/Aitensa/Img2Illum .
[Objective] This paper presents an overview introduction of the VenusAI platform, focusing on its technical updates and sharing the experiences gained since its deployment. The objective is to highlight the platform's advancements, challenges, and valuable insights for other researcher and engineer in the field of AI platform. [Coverage] This paper uses keywords search and citation secondary search to collect papers and information from international computer journals, conferences and open source code warehouse. [Methods] The workflow engine tailored to different processes is designed for streamlining the AI development and enhancing operational efficiency. Then the phased optimization strategy is proposed to address unexpected events and ensure smooth resource allocation operations. Additionally, the disk repair mechanism is utilized to handle disk errors and maintain data integrity. [Experience] We provide detail of the challenges and experiences encountered on the VenusAI platform, aiming to share valuable insights and best practices for AI development. VenusAI's user-friendly toolsets, advanced functionalities, and up-to-date datasets and models make it a leading platform for AI research and development, catering to the diverse needs of both non-IT professionals and advanced researchers. [Conclusion] In conclusion, the VenusAI platform has undergone significant technical updates to improve its performance, efficiency, and stability. The implementation of the workflow engine, operation optimization strategies, and disk repair mechanism has enhanced the platform's capabilities and user experience. The experiences gained from managing VenusAI provide valuable insights for the operation of high-performance computing cluster-based AI platforms. The experience learned and challenges overcome contribute to the continuous improvement and innovation of VenusAI and similar platforms in the future.
The prevalence of personal mobile devices makes iris authentication being more and more popular. Accurate iris segmentation is critical for authentication. However, it is very challenging, due to iris images captured by mobile and handheld devices may exhibit occlusion, low resolution, blur, unusual glint, ghost effect, and off-angles. Moreover, mobile devices may be equipped with visible light cameras rather than near-infrared (NIR) light cameras, which makes iris segmentation susceptible to the noise of visible light. We propose an accurate iris image segmentation approach, which takes advantages of both Conditional Generative Adversarial Network (CGAN) and geometry-based optimization. First, we design a CGAN which force the generator to produce better segmentation corresponds to the original image, a comparatively accurate prediction of iris region can be obtained. Second, a series of geometry-based optimization schemes is introduced to refine the prediction results, where elliptical Hough transform and boundary piecewise fitting are performed on the inner and outer boundary of predicted iris regions, respectively. We performed experiments on three non-ideal iris datasets of visible light and NIR environments. The segmentation accuracy is evaluated using error rate, intersection over union and F-score. Experimental results demonstrate that the proposed approach provides significant performance improvements comparing with the state-of-art methods, OSIRIS and IrisSeg.
Simultaneous localization and mapping (SLAM) is the task to estimate agent's ego-motion in the map and reconstruct the 3D geometric of an unknown environment in parallel. Although many SLAM algorithms have been proposed in the past decades, few efforts have been devoted to conducting accurate real-time dense SLAM on resource- and computation-constrained platforms. In this paper, we leverage a shared binary neural network (BNN) architecture to learn robust feature descriptors for depth estimation and pose estimation modules simultaneously, which not only improves the system's accuracy, but also reduces the computation cost. Also, we propose several optimization strategies targeting feature extraction, feature aggregation as well as feature matching, and to accelerate them on embedded platform. Experimental results demonstrate that our design maintains accurate real-time pose estimation while yielding high-quality dense 3D maps. Our demo video is available at https://github.com/CICAIsubmission/CICAI2023.
Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers in the word. However, the diverse microenvironment, unclear boundaries, integrity destruction inter the slices, and enormous individual differences of tumors pose tremendous challenges to the segmentation process. To address these challenges, we proposed a physicalspiral dual-domain network (PSDD-Net) that combines the advantages of the spiral domain and the physical domain. First of all, the physical domain promotes integral representations of the tumor features, and the spiral domain protrudes the tumor region under CT multi-directions. As a result, the dual-domain framework makes the dual-domain feature simultaneously sent to the network to promote greater attention to the pancreatic region and reduce the interference of redundant background information. Secondly, we also present a multi-scale local-dense net (MSLD-Net) in the physical domain which contains local-channel dense block (LCDB) and multi-scale semantic feature extraction (MSSFE) module. The MSLD-Net grasps more multi-scale geometric information of the tumors and facilitates feature map fusion. Thirdly, a cross-domain aggregation (CDA) module is designed to interact bridging the two domains to interleave and integrate dual-domain complementary visual information. The extensive experiments on the clinical dataset show that our method obtained the DSC of 76.00% in abdominal CT, which outperformed the other state-of-the-art on pancreatic cancer segmentation results and demonstrated strong potential for clinical applications.
Salient object detection in optical remote sensing images aims to detect attractive objects from optical remote sensing images, providing important prior information for many remote sensing tasks, which have received more and more attention in recent years. The existing convolutional neural network-based salient object detection networks mostly rely on pixel-level labelling. Although their detection accuracy is high, annotation cost for the data is high. In addition, it is always a difficult problem that the scales of salient objects in optical remote sensing images change significantly. To address these problems, a new weakly supervised salient object detection method for optical remote sensing images is proposed. Specifically, firstly, we introduce image-level labelling as the weakly supervised information for remote sensing image salient object detection, obtaining pseudo labels to train the saliency detection network. Secondly, we propose the Local Activation Suppression module, including the Discriminative Region Suppression module and Receptive Field Block, which can effectively spread the high response region of the object to the neighbouring low response region, improving the quality of large objects pseudo labels. Finally, the Adaptive Fusion module is proposed to raise the accuracy of pseudo labels of large and small objects, which aims to reduce the noise caused by small objects. Many experiments on a public dataset show that the proposed method is better than the existing weakly supervised learning methods for salient object detection, with better detection accuracy.
