Compact Feature Representation Research Articles

Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation

Semi-supervised learning (SSL), as one of the dominant methods, aims at leveraging the unlabeled data to deal with the annotation dilemma of supervised learning, which has attracted much attentions in the medical image segmentation. Most of the existing approaches leverage a unitary network by convolutional neural networks (CNNs) with compulsory consistency of the predictions through small perturbations applied to inputs or models. The penalties of such a learning paradigm are that (1) CNN-based models place severe limitations on global learning; (2) rich and diverse class-level distributions are inhibited. In this paper, we present a novel CNN-Transformer learning framework in the manifold space for semi-supervised medical image segmentation. First, at intra-student level, we propose a novel class-wise consistency loss to facilitate the learning of both discriminative and compact target feature representations. Then, at inter-student level, we align the CNN and Transformer features using a prototype-based optimal transport method. Extensive experiments show that our method outperforms previous state-of-the-art methods on three public medical image segmentation benchmarks.

GBP: Graph convolutional network embedded in bilinear pooling for fine-grained encoding

In fine-grained recognition, classical high-order coding has inherent contradiction between visual burstiness and feature redundancy, the core of which is the inherent instability of high-order features. Existing methods mainly use EIG and SVD decomposition to maintain feature stability, but this process increases feature redundancy. To address this problem, this paper proposes a Graph Bilinear Pooling (GBP) model to obtain stable fine-grained features through the effective aggregation ability of graph networks. GBP avoids explicit feature decomposition and reconciles the contradiction between visual burstiness and feature redundancy. First, GBP transforms images into a graph spectrum through feature correlation measurement. Then, an improved multi-head graph convolution structure is proposed by using Graph Isomorphism Networks (GIN) to realize feature aggregation. Finally, bilinear pooling operations are performed between graph convolution feature maps and original feature maps to obtain more compact and stable fine-grained feature representations. Experiments on CUB, Cars, and Aircrafts datasets demonstrate that the accuracy of the proposed method is 87.8 %, 93.5 %, and 89.6 % respectively, with a feature representation of 2048 dimensions. Compared to the baseline, the feature number is only 25 % of the baseline model, and the accuracy is increased by 2.6 %, 1.7 %, and 1.3 % respectively. These results demonstrate the effectiveness of graph neural network embedding in improving feature stability.

Prior based Pyramid Residual Clique Network for human body image super-resolution

Recent research in the analysis of human images, such as human parsing and pose estimation, usually requires input images to have a sufficiently high resolution. However, small images of people are commonly encountered in our daily lives, particularly in surveillance applications. This paper aims to ultra-resolve a tiny person image to its high-resolution counterpart by learning effective feature representations and exploiting useful human body prior knowledge. First, we propose the Residual Clique Block (RCB) to fully exploit compact feature representations for image Super-Resolution (SR). Second, a series of RCBs are cascaded in a coarse-to-fine manner to construct the Pyramid Residual Clique Network (PRCN), which simultaneously reconstructs multiple SR results (e.g. 2×, 4×, and 8×) in one feed-forward pass. Third, we utilize the human parsing map as the shape prior, and the high-frequency sub-bands of Uniform Discrete Curvelet Transform (UDCT) as the texture prior to enhance the details of reconstructed human body image. Experimental results demonstrate that our proposed method achieves state-of-the-art performance with superior visual quality and PSNR/SSIM scores. Moreover, we show that our results can considerably enhance the performance of human parsing and pose estimation tasks.

SparseVoxNet: 3-D Object Recognition With Sparsely Aggregation of 3-D Dense Blocks.

Automatic recognition of 3-D objects in a 3-D model by convolutional neural network (CNN) methods has been successfully applied to various tasks, e.g., robotics and augmented reality. Three-dimensional object recognition is mainly performed by analyzing the object using multi-view images, depth images, graphs, or volumetric data. In some cases, using volumetric data provides the most promising results. However, existing recognition techniques on volumetric data have many drawbacks, such as losing object details on converting points to voxels and the large size of the input volume data that leads to substantial 3-D CNNs. Using point clouds could also provide very promising results; however, point-cloud-based methods typically need sparse data entry and time-consuming training stages. Thus, using volumetric could be a more efficient and flexible recognizer for our special case in the School of Medicine, Shanghai Jiao Tong University. In this article, we propose a novel solution to 3-D object recognition from volumetric data using a combination of three compact CNN models, low-cost SparseNet, and feature representation technique. We achieve an optimized network by estimating extra geometrical information comprising the surface normal and curvature into two separated neural networks. These two models provide supplementary information to each voxel data that consequently improve the results. The primary network model takes advantage of all the predicted features and uses these features in Random Forest (RF) for recognition purposes. Our method outperforms other methods in training speed in our experiments and provides an accurate result as good as the state-of-the-art.

Open set recognition fault diagnosis framework based on convolutional prototype learning network for nuclear power plants

As an Open Set Recognition (OSR) problem, nuclear power plant fault diagnosis requires not only the correct classification of known faults, but also the effective identification of unknown faults, which cannot be satisfied by most previously developed models. In this study, a novel nuclear power plant OSR fault diagnosis framework based on Convolutional Prototype Learning (CPL) is proposed, in which CPL is used to extract discriminative fault features from raw nuclear power plant data. Besides, two classification methods, One-Class Support Vector Machine (OCSVM) method and Prototype Matching by Distance (PMD) method, are developed to complete the fault diagnosis of open space for the framework. To verify the feasibility and effectiveness of the proposed OSR framework, numerical experiments on 24 OSR tasks with high-dimensional and strong-nonlinear complex nuclear power plant simulation data are conducted, and the OSR performance evaluate by normalized accuracy and Youden's index, feature visualization and convergence rates are analyzed. Results show that compared with adopting traditional Convolutional Neural Network (CNN), the proposed CPL-based framework can significantly boost the OSR diagnostic performance on nuclear power plant fault diagnosis tasks, which benefits from the excellent ability of CPL to extract intra-class compact and inter-class separable feature representation.

Fusion of Attention-Based Convolution Neural Network and HOG Features for Static Sign Language Recognition

The deaf and hearing-impaired community expresses their emotions, communicates with society, and enhances the interaction between humans and computers using sign language gestures. This work presents a strategy for efficient feature extraction that uses a combination of two different methods that are the convolutional block attention module (CBAM)-based convolutional neural network (CNN) and standard handcrafted histogram of oriented gradients (HOG) feature descriptor. The proposed framework aims to enhance accuracy by extracting meaningful features and resolving issues like rotation, similar hand orientation, etc. The HOG feature extraction technique provides a compact feature representation that signifies meaningful information about sign gestures. The CBAM attention module is incorporated into the structure of CNN to enhance feature learning using spatial and channel attention mechanisms. Then, the final feature vector is formed by concatenating these features. This feature vector is provided to the classification layers to predict static sign gestures. The proposed approach is validated on two publicly available static Massey American Sign Language (ASL) and Indian Sign Language (ISL) databases. The model’s performance is evaluated using precision, recall, F1-score, and accuracy. Our proposed methodology achieved 99.22% and 99.79% accuracy for the ASL and ISL datasets. The acquired results signify the efficiency of the feature fusion and attention mechanism. Our network performed better in accuracy compared to the earlier studies.

Target-Aware Feature Bottleneck for Real-Time Visual Tracking

Recent Siamese network-based visual tracking approaches have achieved high performance metrics on numerous recent visual tracking benchmarks, where most of these trackers employ a backbone feature extractor network with a prediction head network for classification and regression tasks. However, there has been a constant trend of employing a larger and complex backbone network and prediction head networks for improved performance, where increased computational load can slow down the overall speed of the tracking algorithm. To address the aforementioned issues, we propose a novel target-aware feature bottleneck module for trackers, where the proposed bottleneck can elicit a target-aware feature in order to obtain a compact feature representation from the backbone network for improved speed and robustness. Our lightweight target-aware bottleneck module attends to the feature representation of the target region to elicit scene-specific information and generate feature-wise modulation weights that can adaptively change the importance of each feature. The proposed tracker is evaluated on large-scale visual tracking datasets, GOT-10k and LaSOT, and we achieve real-time speed in terms of computation and obtain improved accuracy over the baseline tracker algorithm with high performance metrics.

Quad-Pol SAR Data Reconstruction from Dual-Pol SAR Mode Based on a Multiscale Feature Aggregation Network

Polarimetric synthetic aperture radar (PolSAR) is widely used in remote sensing applications due to its ability to obtain full-polarization information. Compared to the quad-pol SAR, the dual-pol SAR mode has a wider observation swath and is more common in most SAR systems. The goal of reconstructing quad-pol SAR data from the dual-pol SAR mode is to learn the contextual information of dual-pol SAR images and the relationships among polarimetric channels. This work is dedicated to addressing this issue, and a multiscale feature aggregation network has been established to achieve the reconstruction task. Firstly, multiscale spatial and polarimetric features are extracted from the dual-pol SAR images using the pretrained VGG16 network. Then, a group-attention module (GAM) is designed to progressively fuse the multiscale features extracted by different layers. The fused feature maps are interpolated and aggregated with dual-pol SAR images to form a compact feature representation, which integrates the high- and low-level information of the network. Finally, a three-layer convolutional neural network (CNN) with a 1 × 1 convolutional kernel is employed to establish the mapping relationship between the feature representation and polarimetric covariance matrices. To evaluate the quad-pol SAR data reconstruction performance, both polarimetric target decomposition and terrain classification are adopted. Experimental studies are conducted on the ALOS/PALSAR and UAVSAR datasets. The qualitative and quantitative experimental results demonstrate the superiority of the proposed method. The reconstructed quad-pol SAR data can better sense buildings’ double-bounce scattering changes before and after a disaster. Furthermore, the reconstructed quad-pol SAR data of the proposed method achieve a 97.08% classification accuracy, which is 1.25% higher than that of dual-pol SAR data.

Contrastive deep support vector data description

In comparison with support vector data description (SVDD), deep SVDD (DSVDD) is more suitable for dealing with large-scale data sets. DSVDD uses mapping network to replace the role of kernel mapping in SVDD. Moreover, the objective of DSVDD is to simultaneously learn the optimal connection weights of mapping network and the minimum volume of hypersphere. To further improve the performance of DSVDD for tackling large-scale data sets and obtain the discriminative features of the given samples in a self-supervised learning manner, contrastive DSVDD (CDSVDD) is proposed in this study. In the pre-training phase of CDSVDD, the contrastive loss and the rotation prediction loss are jointly minimized to achieve the optimal feature representations. Furthermore, the learned feature representations are utilized to determine the hypersphere center. In the training phase of CDSVDD, the distances between the obtained feature representations and the hypersphere center together with the contrastive loss are simultaneously minimized to derive the optimal network connection weights, the minimum volume of hypersphere and the optimal feature representations. In addition, CDSVDD can efficiently solve the hypersphere collapse problem of DSVDD. The ablation study on CDSVDD verifies that compared with the case of determining the hypersphere center by the feature representations of the original samples, the hypersphere center determined by the feature representations of the augmented samples makes CDSVDD achieve better hypersphere boundary and more compact feature representations. Experimental results on the four benchmark data sets demonstrate that the proposed CDSVDD acquires better detection performance in comparison with its six pertinent methods.

Learning Compact Features via In-Training Representation Alignment

Deep neural networks (DNNs) for supervised learning can be viewed as a pipeline of the feature extractor (i.e., last hidden layer) and a linear classifier (i.e., output layer) that are trained jointly with stochastic gradient descent (SGD) on the loss function (e.g., cross-entropy). In each epoch, the true gradient of the loss function is estimated using a mini-batch sampled from the training set and model parameters are then updated with the mini-batch gradients. Although the latter provides an unbiased estimation of the former, they are subject to substantial variances derived from the size and number of sampled mini-batches, leading to noisy and jumpy updates. To stabilize such undesirable variance in estimating the true gradients, we propose In-Training Representation Alignment (ITRA) that explicitly aligns feature distributions of two different mini-batches with a matching loss in the SGD training process. We also provide a rigorous analysis of the desirable effects of the matching loss on feature representation learning: (1) extracting compact feature representation; (2) reducing over-adaption on mini-batches via an adaptively weighting mechanism; and (3) accommodating to multi-modalities. Finally, we conduct large-scale experiments on both image and text classifications to demonstrate its superior performance to the strong baselines.

D-Net: Learning for distinctive point clouds by self-attentive point searching and learnable feature fusion

Learning and selecting important points on a point cloud is crucial for point cloud understanding in various applications. Most of early methods selected the important points on 3D shapes by analyzing the intrinsic geometric properties of every single shape, which fails to capture the importance of points that distinguishes a shape from objects of other classes, i.e., the distinction of points. To address this problem, we propose D-Net (Distinctive Network) to learn for distinctive point clouds based on a self-attentive point searching and a learnable feature fusion. Specifically, in the self-attentive point searching, we first learn the distinction score for each point to reveal the distinction distribution of the point cloud. After ranking the learned distinction scores, we group a point cloud into a high distinctive point set and a low distinctive one to enrich the fine-grained point cloud structure. To generate a compact feature representation for each distinctive point set, a stacked self-gated convolution is proposed to extract the distinctive features. Finally, we further introduce a learnable feature fusion mechanism to aggregate multiple distinctive features into a global point cloud representation in a channel-wise aggregation manner. The results also show that the learned distinction distribution of a point cloud is highly consistent with objects of the same class and different from objects of other classes. Extensive experiments on public datasets, including ModelNet and ShapeNet part dataset, demonstrate the ability to learn for distinctive point clouds, which helps to achieve the state-of-the-art performance in some shape understanding applications.

Toward Facial Expression Recognition in the Wild via Noise-Tolerant Network

Facial Expression Recognition (FER) has recently emerged as a crucial area in Human-Computer Interaction (HCI) system for understanding the user’s inner state and intention. However, feature- and label-noise constitute the major challenge for FER in the wild due to the ambiguity of facial expressions worsened by low-quality images. To deal with this problem, in this paper, we propose a simple but effective <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Facial Expression Noise-tolerant Network (FENN)</i> which explores the inter-class correlations for mitigating ambiguity that usually happens between morphologically similar classes. Specifically, FENN leverages a multivariate normal distribution to model such correlations at the final hidden layer of the neural network to suppress the heteroscedastic uncertainty caused by inter-class label noise. Furthermore, the discriminative ability of deep features is weakened by the subtle differences between expressions and the presence of feature noise. FENN utilizes a feature-noise mitigation module to extract compact intra-class feature representations under feature noise while preserving the intrinsic inter-class relationships. We conduct extensive experiments to evaluate the effectiveness of FENN on both original annotated images and synthetic noisy annotated images from RAF-DB, AffectNet, and FERPlus in-the-wild facial expression datasets. The results show that FENN significantly outperforms state-of-the-art FER methods.

Comparison of End-to-End Neural Network Architectures and Data Augmentation Methods for Automatic Infant Motility Assessment Using Wearable Sensors

Infant motility assessment using intelligent wearables is a promising new approach for assessment of infant neurophysiological development, and where efficient signal analysis plays a central role. This study investigates the use of different end-to-end neural network architectures for processing infant motility data from wearable sensors. We focus on the performance and computational burden of alternative sensor encoder and time series modeling modules and their combinations. In addition, we explore the benefits of data augmentation methods in ideal and nonideal recording conditions. The experiments are conducted using a dataset of multisensor movement recordings from 7-month-old infants, as captured by a recently proposed smart jumpsuit for infant motility assessment. Our results indicate that the choice of the encoder module has a major impact on classifier performance. For sensor encoders, the best performance was obtained with parallel two-dimensional convolutions for intrasensor channel fusion with shared weights for all sensors. The results also indicate that a relatively compact feature representation is obtainable for within-sensor feature extraction without a drastic loss to classifier performance. Comparison of time series models revealed that feedforward dilated convolutions with residual and skip connections outperformed all recurrent neural network (RNN)-based models in performance, training time, and training stability. The experiments also indicate that data augmentation improves model robustness in simulated packet loss or sensor dropout scenarios. In particular, signal- and sensor-dropout-based augmentation strategies provided considerable boosts to performance without negatively affecting the baseline performance. Overall, the results provide tangible suggestions on how to optimize end-to-end neural network training for multichannel movement sensor data.

Enzyme Substrate Prediction from Three-Dimensional Feature Representations Using Space-Filling Curves.

Compact and interpretable structural feature representations are required for accurately predicting properties and function of proteins. In this work, we construct and evaluate three-dimensional feature representations of protein structures based on space-filling curves (SFCs). We focus on the problem of enzyme substrate prediction, using two ubiquitous enzyme families as case studies: the short-chain dehydrogenase/reductases (SDRs) and the S-adenosylmethionine-dependent methyltransferases (SAM-MTases). Space-filling curves such as the Hilbert curve and the Morton curve generate a reversible mapping from discretized three-dimensional to one-dimensional representations and thus help to encode three-dimensional molecular structures in a system-independent way and with only a few adjustable parameters. Using three-dimensional structures of SDRs and SAM-MTases generated using AlphaFold2, we assess the performance of the SFC-based feature representations in predictions on a new benchmark database of enzyme classification tasks including their cofactor and substrate selectivity. Gradient-boosted tree classifiers yield binary prediction accuracy of 0.77-0.91 and area under curve (AUC) characteristics of 0.83-0.92 for the classification tasks. We investigate the effects of amino acid encoding, spatial orientation, and (the few) parameters of SFC-based encodings on the accuracy of the predictions. Our results suggest that geometry-based approaches such as SFCs are promising for generating protein structural representations and are complementary to the existing protein feature representations such as evolutionary scale modeling (ESM) sequence embeddings.

DML-PL: Deep metric learning based pseudo-labeling framework for class imbalanced semi-supervised learning

Traditional class imbalanced learning algorithms require training data to be labeled, whereas semi-supervised learning algorithms assume that the class distribution is balanced. However, class imbalance and insufficient labeled data problems often coexist in practical real-world applications. Currently, most existing class-imbalanced semi-supervised learning methods tackle these two problems separately, resulting in the trained model biased towards majority classes that have more data samples. In this study, we propose a deep metric learning based pseudo-labeling (DML-PL) framework that tackles both problems simultaneously for class-imbalanced semi-supervised learning. The proposed DML-PL framework comprises three modules: Deep Metric Learning, Pseudo-Labeling and Network Fine-tuning. An iterative self-training strategy is used to train the model multiple times. For each time of training, Deep Metric Learning trains a deep metric network to learn compact feature representations of labeled and unlabeled data. Pseudo-Labeling then generates reliable pseudo-labels for unlabeled data through labeled data clustering with nearest neighbors selection. Finally, Network Fine-tuning fine-tunes the deep metric network to generate better pseudo-labels in the subsequent training. The training ends when all the unlabeled data are pseudo-labeled. The proposed framework achieved state-of-the-art performance on the long-tailed CIFAR-10, CIFAR-100, and ImageNet127 benchmark datasets compared with baseline models.

Deep Graph Convolutional Quantization Networks for Image Retrieval

To achieve real-time online search, most image retrieval methods aim to learn compact feature representation while keeping their semantic information or intra-class relevance. In this paper, we propose a new compact feature learning method to embed the underlying manifold information from database. It integrates deep convolutional neural network (CNN) and graph convolutional neural networks (GCN) into a unified end-to-end learning framework. In the proposed method, the deep feature extracted by CNN is automatically embedded with the information from its neighbors by GCN, which possesses the ability of exploring the semantic relevance on the database manifold. Since constructing a graph over the whole database costs unaffordable memory, we build a landmark graph as database sketch. The landmark graph contains two kinds of nodes, including codewords and memory bank samples. Given an image, the deep architecture outputs the discriminative feature and its similarity with all the graph nodes. We directly use the indices of the most similar codeword nodes as the compact feature representation. To make the proposed method scalable to large datasets, a multi-graph strategy is adopted to generate compact features with adaptable code length. The experiments on two benchmark datasets demonstrate the effectiveness of the proposed method.

Label Propagation and Contrastive Regularization for Semisupervised Semantic Segmentation of Remote Sensing Images

Remarkable progress based on deep neural networks has been achieved on the semantic segmentation in remote sensing images. However, pixel-level labeling is expensive for remote sensing images. Semi-supervised semantic segmentation becomes an alternative approach to reduce the cost of annotation, and it is crucial to utilize efficiently a large number of unlabeled data. Nevertheless inevitably, there is the unbalanced class distribution between labeled and unlabeled data of remote sensing scene. Existing semi-supervised methods train unlabeled images in isolation from labeled images and only learn reliable pixel pseudo-labels, leading to underutilization of unlabeled images. This article proposes a novel semi-supervised semantic segmentation approach based on label propagation and contrastive regularization for remote sensing images. Specifically, the unlabeled images are augmented by randomly copy-pasting the class regions from labeled images. A prototype feature constraint module is used to enforce the constraint on the pixel features of unlabeled images relying on the prototype features from labeled images, achieving feature alignment on the entire dataset. Furthermore, we present the region contrastive learning module that guides the model to learn feature consistency under different perturbations and compact feature representations over class regions on unlabeled images. Extensive experimental results on multiple remote sensing datasets demonstrate that our proposed approach achieves superior performance compared with state-of-the-art semi-supervised semantic segmentation methods.

Domain Generalization for Face Anti-Spoofing via Negative Data Augmentation

In practical applications, the generalization capability of face anti-spoofing (FAS) models on unseen domains is of paramount importance to adapt to diverse camera sensors, device drift, environmental variation, and unpredictable attack types. Recently, various domain generalization (DG) methods have been developed to improve the generalization capability of FAS models via training on multiple source domains. These DG methods commonly require collecting sufficient real-world attack samples of different attack types for each source domain. This work aims to learn a FAS model without using any real-world attack sample in any source domain but can generalize well to the unseen domain, which can significantly reduce the learning cost. Toward this goal, we draw inspiration from the theoretical error bound of domain generalization to use negative data augmentation instead of real-world attack samples for training. We show that using only a few types of simple synthesized negative samples, e.g., color jitter and color mask, the learned model can achieve competitive performance over state-of-the-art DG methods trained using real-world attack samples. Moreover, a dynamic global common loss and a local contrast loss are proposed to prompt the model to learn a compact and common feature representation for real face samples from different source domains, which can further improve the generalization capability. Experimental results of extensive cross-dataset testing demonstrate that our method can even outperform state-of-the-art DG methods using real-world attack samples for training. The code for reproducing the results of our method is available at https://github.com/WeihangWANG/NDA-FAS.

Centralized Feature Pyramid for Object Detection.

The visual feature pyramid has shown its superiority in both effectiveness and efficiency in a variety of applications. However, current methods overly focus on inter-layer feature interactions while disregarding the importance of intra-layer feature regulation. Despite some attempts to learn a compact intra-layer feature representation with the use of attention mechanisms or vision transformers, they overlook the crucial corner regions that are essential for dense prediction tasks. To address this problem, we propose a Centralized Feature Pyramid (CFP) network for object detection, which is based on a globally explicit centralized feature regulation. Specifically, we first propose a spatial explicit visual center scheme, where a lightweight MLP is used to capture the globally long-range dependencies, and a parallel learnable visual center mechanism is used to capture the local corner regions of the input images. Based on this, we then propose a globally centralized regulation for the commonly-used feature pyramid in a top-down fashion, where the explicit visual center information obtained from the deepest intra-layer feature is used to regulate frontal shallow features. Compared to the existing feature pyramids, CFP not only has the ability to capture the global long-range dependencies but also efficiently obtain an all-round yet discriminative feature representation. Experimental results on the challenging MS-COCO validate that our proposed CFP can achieve consistent performance gains on the state-of-the-art YOLOv5 and YOLOX object detection baselines.

Many Hands Make Light Work: Transferring Knowledge From Auxiliary Tasks for Video-Text Retrieval

The problem of video-text retrieval, which searches videos via natural language descriptions or vice versa, has attracted growing attention due to the explosive scale of videos produced every day. The dominant approaches for this problem follow the pipeline that firstly learn compact feature representations of videos and texts, and then jointly embed them into a common feature space where matched video-text pairs are close and unmatched pairs are far away. However, most of them neither consider the structural similarities among crossmodal samples in a global view, nor leverage useful information from other relevant retrieval processes. We argue that both information have great potential for video-text retrieval. In this paper, we propose to extract useful knowledge from the retrieval process by exploiting structural similarities via Graph Neural Networks (GNNs) and then progressively transfer useful knowledge from relevant retrieval processes in a general-tospecific manner to assist the current retrieval process. Specifically, for the retrieval of the current query, we first construct a sequence of query-graphs whose central queries are chosen from distant to close to the current query. Then we conduct knowledgeguided message passing in each query-graph to exploit regional structural similarities and gather knowledge of different levels from the updated query-graphs with a knowledge-based attention mechanism. Finally, we transfer the extracted useful knowledge from general to specific to assist the current retrieval process. Extensive experimental results show that our model outperforms the state-of-the-arts on four benchmarks.