Multi-scale Deep Learning Research Articles

A Multiscale Deep Learning Method for Quantitative Visualization of Traumatic Hemoperitoneum at CT: Assessment of Feasibility and Comparison with Subjective Categorical Estimation.

To evaluate the feasibility of a multiscale deep learning algorithm for quantitative visualization and measurement of traumatic hemoperitoneum and to compare diagnostic performance for relevant outcomes with categorical estimation. This retrospective, single-institution study included 130 patients (mean age, 38 years; interquartile range, 25-50 years; 79 men) with traumatic hemoperitoneum who underwent CT of the abdomen and pelvis at trauma admission between January 2016 and April 2019. Labeled cases were separated into five combinations of training (80%) and test (20%) sets, and fivefold cross-validation was performed. Dice similarity coefficients (DSCs) were compared with those from a three-dimensional (3D) U-Net and a coarse-to-fine deep learning method. Areas under the receiver operating characteristic curve (AUCs) for a composite outcome, including hemostatic intervention, transfusion, and in-hospital mortality, were compared with consensus categorical assessment by two radiologists. An optimal cutoff was derived by using a radial basis function-based support vector machine. Mean DSC for the multiscale algorithm was 0.61 ± 0.15 (standard deviation) compared with 0.32 ± 0.16 for the 3D U-Net method and 0.52 ± 0.17 for the coarse-to-fine method (P < .0001). Correlation and agreement between automated and manual volumes were excellent (Pearson r = 0.97, intraclass correlation coefficient = 0.93). The algorithm produced intuitive and explainable visual results. AUCs for automated volume measurement and categorical estimation were 0.86 and 0.77, respectively (P = .004). An optimal cutoff of 278.9 mL yielded accuracy of 84%, sensitivity of 82%, specificity of 93%, positive predictive value of 86%, and negative predictive value of 83%. A multiscale deep learning method for traumatic hemoperitoneum quantitative visualization had improved diagnostic performance for predicting hemorrhage-control interventions and mortality compared with subjective volume estimation. Supplemental material is available for this article. © RSNA, 2020.

Deep multi-scale learning for automatic tracking of internal layers of ice in radar data

AbstractIn this study, our goal is to track internal ice layers on the Snow Radar data collected by NASA Operation IceBridge. We examine the application of deep learning methods on radar data gathered from polar regions. Artificial intelligence techniques have displayed impressive success in many practical fields. Deep neural networks owe their success to the availability of massive labeled data. However, in many real-world problems, even when a large dataset is available, deep learning methods have shown less success, due to causes such as lack of a large labeled dataset, presence of noise in the data or missing data. In our radar data, the presence of noise is one of the main obstacles in utilizing popular deep learning methods such as transfer learning. Our experiments show that if the neural network is trained to detect contours of objects in electro-optical imagery, it can only track a low percentage of contours in radar data. Fine-tuning and further training do not provide any better results. However, we show that selecting the right model and training it on the radar imagery from the start yields far better results.

Automatic Drusen Segmentation for Age-Related Macular Degeneration in Fundus Images Using Deep Learning

Drusen are the main aspect of detecting age-related macular degeneration (AMD). Ophthalmologists can evaluate the condition of AMD based on drusen in fundus images. However, in the early stage of AMD, the drusen areas are usually small and vague. This leads to challenges in the drusen segmentation task. Moreover, due to the high-resolution fundus images, it is hard to accurately predict the drusen areas with deep learning models. In this paper, we propose a multi-scale deep learning model for drusen segmentation. By exploiting both local and global information, we can improve the performance, especially in the early stages of AMD cases.

Difficulty-aware hierarchical convolutional neural networks for deformable registration of brain MR images.

The aim of deformable brain image registration is to align anatomical structures, which can potentially vary with large and complex deformations. Anatomical structures vary in size and shape, requiring the registration algorithm to estimate deformation fields at various degrees of complexity. Here, we present a difficulty-aware model based on an attention mechanism to automatically identify hard-to-register regions, allowing better estimation of large complex deformations. The difficulty-aware model is incorporated into a cascaded neural network consisting of three sub-networks to fully leverage both global and local contextual information for effective registration. The first sub-network is trained at the image level to predict a coarse-scale deformation field, which is then used for initializing the subsequent sub-network. The next two sub-networks progressively optimize at the patch level with different resolutions to predict a fine-scale deformation field. Embedding difficulty-aware learning into the hierarchical neural network allows harder patches to be identified in the deeper sub-networks at higher resolutions for refining the deformation field. Experiments conducted on four public datasets validate that our method achieves promising registration accuracy with better preservation of topology, compared with state-of-the-art registration methods.

Multi-scale deep feature learning network with bilateral filtering for SAR image classification

Synthetic aperture radar (SAR) image classification using deep neural network has drawn great attention, which generally requires various layers of deep model for feature learning. However, a deeper neural network will result in overfitting with limited training samples. In this paper, a multi-scale deep feature learning network with bilateral filtering (MDFLN-BF) is proposed for SAR image classification, which aims to extract discriminative features and reduce the requirement of labeled samples. In the proposed framework, MDFLN is proposed to extract features from SAR image on multiple scales, where the SAR image is stratified into different scales and a full convolutional network is utilized to extract features from each scale sub-image. Then, features of multiple scales are classified by multiple softmax classifiers and combined by majority vote algorithm. Further, bilateral filtering is developed to optimize the classification map based on spatial relation, which aims to improve the spatial smoothness. Experiments are tested on three SAR images with different sensors, bands, resolutions, and polarizations in order to prove the generalization ability. It is demonstrated that the proposed MDFLN-BF is able to yield superior results than other related deep networks.

Noncontact super-resolution guided wave array imaging of subwavelength defects using a multiscale deep learning approach

Subwavelength defect imaging using guided waves has been known to be a difficult task mainly due to the diffraction limit and dispersion of guided waves. In this article, we present a noncontact super-resolution guided wave array imaging approach based on deep learning to visualize subwavelength defects in plate-like structures. The proposed approach is a novel hierarchical multiscale imaging approach that combines two distinct fully convolutional networks. The first fully convolutional network, the global detection network, globally detects subwavelength defects in a raw low-resolution guided wave beamforming image. Then, the subsequent second fully convolutional network, the local super-resolution network, locally resolves subwavelength-scale fine structural details of the detected defects. We conduct a series of numerical simulations and laboratory-scale experiments using a noncontact guided wave array enabled by a scanning laser Doppler vibrometer on aluminate plates with various subwavelength defects. The results demonstrate that the proposed super-resolution guided wave array imaging approach not only locates subwavelength defects but also visualizes super-resolution fine structural details of these defects, thus enabling further estimation of the size and shape of the detected subwavelength defects. We discuss several key aspects of the performance of our approach, compare with an existing super-resolution algorithm, and make recommendations for its successful implementations.

Multi-scale deep intra-class transfer learning for bearing fault diagnosis

The tremendous success of deep learning in machine fault diagnosis is dependent on the hypothesis that training and test datasets are subordinated to the same distribution. This subordination is difficult to meet in practical scenarios of industrial applications. On the one hand, the working conditions of rotating machinery can change easily. On the other hand, vibration data and labels are difficult to obtain to train a specific model for each working condition. In this study, we solve these problems by constructing a novel deep transfer learning model called multi-scale deep intra-class adaptation network, which first uses the modified ResNet-50 to extract low-level features and then constructs a multiple scale feature learner to analyze these low-level features at multiple scales and obtain high-level features as input for the classifier. Pseudo labels are then computed to shorten the conditional distribution distance of vibration data collected under different working loads for intra-class adaptation. The proposed method is validated using two datasets to recognize the bearing normal state, the inner race, the ball and outer race faults, and their fault degrees under four different working loads. The high-precision diagnosis results of 24 transfer learning experiments reveal the reliability and generalizability of the constructed model.

A tool wear monitoring and prediction system based on multiscale deep learning models and fog computing

Tool condition monitoring (TCM) during the manufacturing process is of great significance for ensuring product quality and plays an important role in intelligent manufacturing. Current TCM systems deployed in the local device or cloud computing environment unable meet the requirements of low response latency and high accuracy at the same time. The emerging fog computing provides new solutions for the above problem. This paper presents a tool wear monitoring and prediction (TWMP) system based on deep learning models and fog computing. In order to improve monitoring and prediction accuracy, we propose a multiscale convolutional long short-term memory model (MCLSTM) to complete the tool wear monitoring task and a bi-directional LSTM model (BiLSTM) to complete the tool wear prediction task. To reduce the response latency of the TWMP system, we deploy the MCLSTM model and the BiLSTM model in a fog computing architecture. The fog computing architecture consists of an edge computing layer, a fog computing layer, and a cloud computing layer. The edge computing layer undertakes real-time signal collection task. The fog computing layer undertakes real-time tool wear monitoring task. The cloud computing layer with powerful computing resources undertakes intensive computing and latency-insensitive tasks such as data storage, tool wear prediction, and model training. A twist drill wear monitoring and prediction experiment is conducted to test the performance of the proposed system in terms of accuracy, response time, and network bandwidth consumption.

Identifying Informal Settlements Using Contourlet Assisted Deep Learning.

As the global urban population grows due to the influx of migrants from rural areas, many cities in developing countries face the emergence and proliferation of unplanned and informal settlements. However, even though the rise of unplanned development influences planning and management of residential land-use, reliable and detailed information about these areas is often scarce. While formal settlements in urban areas are easily mapped due to their distinct features, this does not hold true for informal settlements because of their microstructure, instability, and variability of shape and texture. Therefore, detecting and mapping these areas remains a challenging task. This research will contribute to the development of tools to identify such informal built-up areas by using an integrated approach of multiscale deep learning. The authors propose a composite architecture for semantic segmentation using the U-net architecture aided by information obtained from a multiscale contourlet transform. This work also analyzes the effects of wavelet and contourlet decompositions in the U-net architecture. The performance was evaluated in terms of precision, recall, F-score, mean intersection over union, and overall accuracy. It was found that the proposed method has better class-discriminating power as compared to existing methods and has an overall classification accuracy of 94.9–95.7%.

Object-oriented multiscale deep features for hyperspectral image classification

ABSTRACT The classification of hyperspectral images (HSIs) is one of the most popular topics in the remote sensing community. Numerous feature extraction methods have been proposed to improve the classification accuracy of HSIs. Recently, deep features extracted by convolution neural network (CNN) have been introduced into the classification process of HSIs. Due to the nonlinear and invariant advantages of the features, CNN methods provide a powerful tool for representing geographic objects and classifying HSIs. However, traditional deep features only extracted at pixel-level and often neglect multiscale characteristics of geographic objects. In this study, a new deep feature extraction method is proposed, which takes advantage of multi-scale object analysis and the CNN model. Firstly, multiscale image objects are obtained by the multiscale segmentation algorithm and multiscale low-level features of objects are extracted. Secondly, the CNN is devoted to obtain deep features from low-level object features at each scale, respectively. Thirdly, the obtained deep features at all scales are stacked and fed to one fully connected layer to extract the multiscale deep learning features for classification. Finally, the logistic regression classifier is applied to hyperspectral image (HSI) classification based on object-oriented multiscale deep features. The proposed method was carried out on three widely used hyperspectral data sets: University of Pavia, Salinas, and Washington DC. The results reveal that the proposed method provides better results than other state-of-the-art methods.

Deep point embedding for urban classification using ALS point clouds: A new perspective from local to global

Semantic interpretation of the 3D scene is one of the most challenging problems in point cloud processing, which also deems as an essential task in a wide variety of point cloud applications. The core task of semantic interpretation is semantic labeling, namely, obtaining a unique semantic label for each point in the point cloud. Despite several reported approaches, semantic labeling continues to be a challenge owing to the complexity of scenes, objects of various scales, and the non-homogeneity of unevenly distributed points. In this paper, we propose a novel method for obtaining semantic labels of airborne laser scanning (ALS) point clouds involving the embedding of local context information for each point with multi-scale deep learning, nonlinear manifold learning for feature dimension reduction, and global graph-based optimization for refining the classification results. Specifically, we address the tasks of learning discriminative features and global labeling smoothing. The key contribution of our study is threefold. First, a hierarchical data augmentation strategy is applied to enhance the learning of deep features based on the PointNet++ network and simultaneously eliminate the artifacts caused by division and sampling while dealing with large-scale datasets. Subsequently, the learned hierarchical deep features are globally optimized and embedded into a low-dimensional space with a nonlinear manifold-based joint learning method with the removal of redundant and disturbing information. Finally, a graph-structured optimization based on the Markov random fields algorithm is performed to achieve global optimization of the initial classification results that are obtained using the embedded deep features by constructing a weighted indirect graph and solving the optimization problem with graph-cuts. We conducted thorough experiments on ALS point cloud datasets to assess the performance of our framework. Results indicate that compared to other commonly used advanced classification methods, our method can achieve high classification accuracy. The overall accuracy (OA) of our approach on the ISPRS benchmark dataset can scale up to 83.2% for classifying nine semantic classes, thereby outperforming other compared point-based strategies. Additionally, we evaluated our framework on a selected portion of the AHN3 dataset, which provided OA up to 91.2%.

Validation of a fully automated liver segmentation algorithm using multi-scale deep reinforcement learning and comparison versus manual segmentation

PurposeTo evaluate the performance of an artificial intelligence (AI) based software solution tested on liver volumetric analyses and to compare the results to the manual contour segmentation. Materials and methodsWe retrospectively obtained 462 multiphasic CT datasets with six series for each patient: three different contrast phases and two slice thickness reconstructions (1.5/5 mm), totaling 2772 series. AI-based liver volumes were determined using multi-scale deep-reinforcement learning for 3D body markers detection and 3D structure segmentation. The algorithm was trained for liver volumetry on approximately 5000 datasets. We computed the absolute error of each automatically- and manually-derived volume relative to the mean manual volume. The mean processing time/dataset and method was recorded. Variations of liver volumes were compared using univariate generalized linear model analyses. A subgroup of 60 datasets was manually segmented by three radiologists, with a further subgroup of 20 segmented three times by each, to compare the automatically-derived results with the ground-truth. ResultsThe mean absolute error of the automatically-derived measurement was 44.3 mL (representing 2.37 % of the averaged liver volumes). The liver volume was neither dependent on the contrast phase (p = 0.697), nor on the slice thickness (p = 0.446). The mean processing time/dataset with the algorithm was 9.94 s (sec) compared to manual segmentation with 219.34 s. We found an excellent agreement between both approaches with an ICC value of 0.996. ConclusionThe results of our study demonstrate that AI-powered fully automated liver volumetric analyses can be done with excellent accuracy, reproducibility, robustness, speed and agreement with the manual segmentation.

Multi-Scale deep learning framework for cochlea localization, segmentation and analysis on clinical ultra-high-resolution CT images

Background and objectivePerforming patient-specific, pre-operative cochlea CT-based measurements could be helpful to positively affect the outcome of cochlear surgery in terms of intracochlear trauma and loss of residual hearing. Therefore, we propose a method to automatically segment and measure the human cochlea in clinical ultra-high-resolution (UHR) CT images, and investigate differences in cochlea size for personalized implant planning. Methods123 temporal bone CT scans were acquired with two UHR-CT scanners, and used to develop and validate a deep learning-based system for automated cochlea segmentation and measurement. The segmentation algorithm is composed of two major steps (detection and pixel-wise classification) in cascade, and aims at combining the results of a multi-scale computer-aided detection scheme with a U-Net-like architecture for pixelwise classification. The segmentation results were used as an input to the measurement algorithm, which provides automatic cochlear measurements (volume, basal diameter, and cochlear duct length (CDL)) through the combined use of convolutional neural networks and thinning algorithms. Automatic segmentation was validated against manual annotation, by the means of Dice similarity, Boundary-F1 (BF) score, and maximum and average Hausdorff distances, while measurement errors were calculated between the automatic results and the corresponding manually obtained ground truth on a per-patient basis. Finally, the developed system was used to investigate the differences in cochlea size within our patient cohort, to relate the measurement errors to the actual variation in cochlear size across different patients. ResultsAutomatic segmentation resulted in a Dice of 0.90 ± 0.03, BF score of 0.95 ± 0.03, and maximum and average Hausdorff distance of 3.05 ± 0.39 and 0.32 ± 0.07 against manual annotation. Automatic cochlear measurements resulted in errors of 8.4% (volume), 5.5% (CDL), 7.8% (basal diameter). The cochlea size varied broadly, ranging between 0.10 and 0.28 ml (volume), 1.3 and 2.5 mm (basal diameter), and 27.7 and 40.1 mm (CDL). ConclusionsThe proposed algorithm could successfully segment and analyze the cochlea on UHR-CT images, resulting in accurate measurements of cochlear anatomy. Given the wide variation in cochlear size found in our patient cohort, it may find application as a pre-operative tool in cochlear implant surgery, potentially helping elaborate personalized treatment strategies based on patient-specific, image-based anatomical measurements.

Image Interpolation Using Multi-scale Attention-aware Inception Network.

A new multi-scale deep learning (MDL) framework is proposed and exploited for conducting image interpolation in this paper. The core of the framework is a seeding network that needs to be designed for the targeted task. For image interpolation, a novel attention-aware inception network (AIN) is developed as the seeding network; it has two key stages: 1) feature extraction based on the low-resolution input image; and 2) feature-to-image mapping to enlarge image's size or resolution. Note that the designed seeding network, AIN, needs to be trained with a matched training dataset at each scale. For that, multi-scale image patches are generated using our proposed pyramid cut, which outperforms the conventional image pyramid method by completely avoiding aliasing issue. After training, the trained AINs are then combined for processing the input image in the testing stage. Extensive experimental simulation results obtained from seven image datasets (comprising 359 images in total) have clearly shown that the proposed MAIN consistently delivers highly accurate interpolated images.

A Two-Stream Multiscale Deep Learning Architecture for Pan-Sharpening

Pan-sharpening, which fuses the high-resolution panchromatic (PAN) image and the low-resolution multispectral image (MSI), is a hot topic in remote sensing. Recently, deep learning technology has been successfully applied in pan-sharpening. However, the existing methods ignore that the MSI and PAN image are at different resolutions and use the same networks to extract features of the two images. To address this problem, we propose a two-stream deep learning architecture, called coupled multiscale convolutional neural network, for pan-sharpening. The proposed network has three components, feature extraction subnetworks, fusion layer, and super-resolution subnetwork. In the feature extraction subnetworks, two subnetworks are used to extract the features of the MSI and PAN image separately. Different sizes of convolutional kernels are used in the first layers due to the different spatial resolutions. Thus, the source images are mapped to the similar scale. Then a multiscale asymmetric convolution factorization is used to extract features at different scales. In the fusion layer, the two feature extraction subnetworks are coupled. Features at the same scale are first summed, and then the features of all scales are concatenated as one feature map. At last, a super-resolution subnetwork is used to generate the high-resolution MSI. Experimental results on both synthetic and real data sets demonstrate that the proposed method outperforms the other state-of-the-art pan-sharpening methods.

EMMCNN: An ETPS-Based Multi-Scale and Multi-Feature Method Using CNN for High Spatial Resolution Image Land-Cover Classification

Land-cover information is significant for land-use planning, urban management, and environment monitoring. This paper presented a novel extended topology-preserving segmentation (ETPS)-based multi-scale and multi-feature method using the convolutional neural network (EMMCNN) for high spatial resolution (HSR) image land-cover classification. The EMMCNN first segmented the images into superpixels using the ETPS algorithm with false-color composition and enhancement and built parallel convolutional neural networks (CNNs) with dense connections for superpixel multi-scale deep feature learning. Then, the multi-resolution segmentation (MRS) object hand-delineated features were extracted and mapped to superpixels for complementary multi-segmentation and multi-type representation. Finally, a hybrid network was designed to consist of 1-dimension CNN and multi-layer perception (MLP) with channel-wise stacking and attention-based weighting for adaptive feature fusion and comprehensive classification. Experimental results on four real HSR GaoFen-2 datasets demonstrated the superiority of the proposed EMMCNN over several well-known classification methods in terms of accuracy and consistency, with overall accuracy averagely improved by 1.74% to 19.35% for testing images and 1.06% to 8.78% for validating images. It was found that the solution combining an appropriate number of larger scales and multi-type features is recommended for better performance. Efficient superpixel segmentation, networks with strong learning ability, optimized multi-scale and multi-feature solution, and adaptive attention-based feature fusion were key points for improving HSR image land-cover classification in this study.

Topology Prediction Improvement of α-helical Transmembrane Proteins Through Helix-tail Modeling and Multiscale Deep Learning Fusion

Transmembrane proteins (TMPs) play important roles in many biological processes, such as cell recognition and communication. Their structures are crucial for revealing complex functions but are hard to obtain. A variety of computational algorithms have been proposed to fill the gap by predicting structures from primary sequences. In this study, we mainly focus on α-helical TMP and develop a multiscale deep learning pipeline, MemBrain 3.0, to improve topology prediction. This new protocol includes two submodules. The first module is transmembrane helix (TMH) prediction, which features the capability of accurately predicting TMH with the tail part through the incorporation of tail modeling. The prediction engine contains a multiscale deep learning model and a dynamic threshold strategy. The deep learning model is comprised of a small-scale residue-based residual neural network and a large-scale entire-sequence-based residual neural network. Dynamic threshold strategy is designed to binarize the raw prediction scores and solve the under-split problem. The second module is orientation prediction, which consists of a support vector machine (SVM) classifier and a new Max-Min assignment (MMA) strategy. One typical merit of MemBrain 3.0 is the decision mode composed of the dynamic threshold strategy and the MMA strategy, which makes it more effective for hard TMHs, such as half-TMH, back-to-back TMH, and long-TMH. Systematic experiments have demonstrated the efficacy of the new model, which is available at: www.csbio.sjtu.edu.cn/bioinf/MemBrain/.

MDFN: Multi-scale deep feature learning network for object detection

This paper proposes an innovative object detector by leveraging deep features learned in high-level layers. Compared with features produced in earlier layers, the deep features are better at expressing semantic and contextual information. The proposed deep feature learning scheme shifts the focus from concrete features with details to abstract ones with semantic information. It considers not only individual objects and local contexts but also their relationships by building a multi-scale deep feature learning network (MDFN). MDFN efficiently detects the objects by introducing information square and cubic inception modules into the high-level layers, which employs parameter-sharing to enhance the computational efficiency. MDFN provides a multi-scale object detector by integrating multi-box, multi-scale and multi-level technologies. Although MDFN employs a simple framework with a relatively small base network (VGG-16), it achieves better or competitive detection results than those with a macro hierarchical structure that is either very deep or very wide for stronger ability of feature extraction. The proposed technique is evaluated extensively on KITTI, PASCAL VOC, and COCO datasets, which achieves the best results on KITTI and leading performance on PASCAL VOC and COCO. This study reveals that deep features provide prominent semantic information and a variety of contextual contents, which contribute to its superior performance in detecting small or occluded objects. In addition, the MDFN model is computationally efficient, making a good trade-off between the accuracy and speed.

Multi-Scale Deep Residual Learning-Based Single Image Haze Removal via Image Decomposition.

Images/videos captured from outdoor visual devices are usually degraded by turbid media, such as haze, smoke, fog, rain, and snow. Haze is the most common one in outdoor scenes due to the atmosphere conditions. In this paper, a novel deep learning-based architecture (denoted by MSRL-DehazeNet) for single image haze removal relying on multi-scale residual learning (MSRL) and image decomposition is proposed. Instead of learning an end-to-end mapping between each pair of hazy image and its corresponding haze-free one adopted by most existing learningbased approaches, we reformulate the problem as restoration of the image base component. Based on the decomposition of a hazy image into the base and the detail components, haze removal (or dehazing) can be achieved by both of our multi-scale deep residual learning and our simplified U-Net learning only for mapping between hazy and haze-free base components, while the detail component is further enhanced via the other learned convolutional neural network (CNN). Moreover, benefited by the basic building block of our deep residual CNN architecture and our simplified UNet structure, the feature maps (produced by extracting structural and statistical features), and each previous layer can be fully preserved and fed into the next layer. Therefore, possible color distortion in the recovered image would be avoided. As a result, the final haze-removed (or dehazed) image is obtained by integrating the haze-removed base and the enhanced detail image components. Experimental results have demonstrated good effectiveness of the proposed framework, compared with state-ofthe-art approaches.

Leader-Based Multi-Scale Attention Deep Architecture for Person Re-Identification.

Person re-identification (re-id) aims to match people across non-overlapping camera views in a public space. This is a challenging problem because the people captured in surveillance videos often wear similar clothing. Consequently, the differences in their appearance are typically subtle and only detectable at particular locations and scales. In this paper, we propose a deep re-id network (MuDeep) that is composed of two novel types of layers - a multi-scale deep learning layer, and a leader-based attention learning layer. Specifically, the former learns deep discriminative feature representations at different scales, while the latter utilizes the information from multiple scales to lead and determine the optimal weightings for each scale. The importance of different spatial locations for extracting discriminative features is learned explicitly via our leader-based attention learning layer. Extensive experiments are carried out to demonstrate that the proposed MuDeep outperforms the state-of-the-art on a number of benchmarks and has a better generalization ability under a domain generalization setting.