Improve Retrieval Performance Research Articles

A novel re-ranking architecture for patent search

Patent search presents unique challenges due to the intricate structure and specialized terminology embedded in patent documents. While neural models have been successfully applied in various information retrieval (IR) tasks, these inherent complexities have hindered their effectiveness in patent search. To address these challenges, we propose a novel re-ranking architecture that effectively handles long, structured patent documents and leverages AI models to interpolate lexical and semantic signals of relevance. Additionally, the architecture incorporates query-specific weights for the final re-ranking process. To address partial relevance between patent sections our method effectively models the relevance relationships between different sections of patent documents. We calculate lexical and semantic signals of relevance from each document section and feed them as input features to AI models that estimate a combined relevance score. Finally, we compute query-specific weights to determine the relative contributions of lexical and semantic relevance for the final re-ranking. Extensive experiments on the CLEF-IP dataset demonstrate that our method outperforms several baselines, achieving substantial and statistically significant improvements in retrieval performance. We further assess the adaptability of our method using the MSMARCO dataset, where it exhibits limited performance, indicating its suitability for domain-specific patent research.

Parallel Hierarchical Multi-View Feature Fusion Based on Canonical Correlation Analysis for Mammogram Retrieval

Due to the diversity of image sources, content-based multi-source image fusion and retrieval have shown promising capabilities in computer vision tasks, and especially when applied in Computer-Aided Diagnosis (CAD) to automate and improve the accuracy of medical image analysis. The combination of computer vision and CAD systems has the potential to revolutionize healthcare by augmenting the expertise of clinicians, improving overall diagnostic accuracy and helping experts in the clinical decision-making process by classifying and retrieving similar annotated clinical images to a given query. In the context of multi-view mammography interpretation, the concept of multi-view feature fusion has recently been studied to improve retrieval performance while effectively guaranteeing the complementarity of both MLO and CC views. However, conventional multi-view feature fusion makes descriptors long and lacks to take into consideration the relationship between descriptors. To deal with this issue, we propose two hierarchical multi-view feature fusion methods, for multi-view mammogram retrieval, based on the Canonical Correlation Analysis (CCA), which is the most commonly used multivariate parametric test. In fact, we have adapted CCA to determine the relationship between two descriptors by processing latent correlation factors. Moreover, after extracting descriptors for each view, a comparative study of texture and shape fusion descriptors is proposed in order to identify the more discriminative features for multi-view mammogram retrieval. Then, a query-dependent distance metric preserving both visual resemblance and semantic similarity is carried out to dynamically determine the more appropriate distance measure for each query image. Extensive experiments on the challenging Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) have demonstrated the effectiveness of the proposed hierarchical multi-view feature fusion for mammogram retrieval, which outperforms the performance achieved either by conventional fused information or by single view information. To improve the transparency of our paper, the source code of the proposed method and the related dataset (including readme files) are publicly accessible through the following GitHub link: https://github.com/ABDERRAHIMMAR/Multi-View-Feature-Fusion-for-Mammogram-Retrieval . This open-access resource empowers researchers and practitioners to delve deeper into our methodology, fostering collaboration and advancements in the field of computer-aided diagnosis and medical image analysis.

SCA-PVNet: Self-and-cross attention based aggregation of point cloud and multi-view for 3D object retrieval

To address 3D object retrieval, substantial efforts have been made to generate highly discriminative descriptors for 3D objects represented by a single modality, such as voxels, point clouds, or multiview images. It is promising to leverage complementary information from multimodal representations of 3D objects to further improve retrieval performance. However, multimodal 3D object retrieval has rarely been developed or analyzed for large-scale datasets. In this paper, we propose a self-and-cross-attention-based aggregation of point clouds and multiview images (SCA-PVNet) for 3D object retrieval. With deep features extracted from point clouds and multi-view images, we design two types of feature aggregation modules, namely the in-modality aggregation module (IMAM) and the cross-modality aggregation module (CMAM ), for effective feature fusion. IMAM leverages a self-attention mechanism to aggregate multiview features, whereas CMAM exploits a cross-attention mechanism to interact with point-cloud and multiview features. The final descriptor of a 3D object for object retrieval can be obtained by concatenating the aggregated feature outputs of both modules. Extensive experiments and analyses were conducted on four datasets ranging from small to large scales to demonstrate the superiority of the proposed SCA-PVNet over state-of-the-art methods. In addition to achieving state-of-the-art retrieval performance, our method is more robust in challenging scenarios when views or points are missing during inference.

Text-based person search via cross-modal alignment learning

Text-based person search aims to use text descriptions to search for corresponding person images. However, due to the obvious pattern differences in image and text modalities, it is still a challenging problem to align the two modalities. Most existing approaches only consider semantic alignment within a global context or partial parts, lacking consideration of how to match image and text in terms of differences in model information. Therefore, in this paper, we propose an efficient Modality-Aligned Person Search network (MAPS) to address this problem. First, we suppress image-specific information by image feature style normalization to achieve modality knowledge alignment and reduce information differences between text and image. Second, we design a multi-granularity modal feature fusion and optimization method to enrich the modal features. To address the problem of useless and redundant information in the multi-granularity fused features, we propose a Multi-granularity Feature Self-optimization Module (MFSM) to adaptively adjust the corresponding contributions of different granularities in the fused features of the two modalities. Finally, to address the problem of information inconsistency in the training and inference stages, we propose a Cross-instance Feature Alignment (CFA) to help the network enhance category-level generalization ability and improve retrieval performance. Extensive experiments demonstrate that our MAPS achieves state-of-the-art performance on all text-based person search datasets, and significantly outperforms other existing methods.

Relevance Feedback with Brain Signals

The Relevance Feedback (RF) process relies on accurate and real-time relevance estimation of feedback documents to improve retrieval performance. Since collecting explicit relevance annotations imposes an extra burden on the user, extensive studies have explored using pseudo-relevance signals and implicit feedback signals as substitutes. However, such signals are indirect indicators of relevance and suffer from complex search scenarios where user interactions are absent or biased. Recently, the advances in portable and high-precision brain-computer interface (BCI) devices have shown the possibility to monitor user’s brain activities during search process. Brain signals can directly reflect user’s psychological responses to search results and thus it can act as additional and unbiased RF signals. To explore the effectiveness of brain signals in the context of RF, we propose a novel RF framework that combines BCI-based RF with pseudo-relevance signals and implicit signals to improve the performance of document re-ranking. The experimental results on the user study dataset show that incorporating brain signals leads to significant performance improvement in our RF framework. Besides, we observe that brain signals perform particularly well in several hard search scenarios, especially when implicit signals as feedback are missing or noisy. This reveals when and how to exploit brain signals in the context of RF.

A set of novel HTML document quality features for Web information retrieval: Including applications to learning to rank for information retrieval

The past work on Information Retrieval (IR) targeting web document collections shows that incorporating a measure that measures the quality of web documents, or rather the document prior (e.g., PageRank), into an IR system improves the retrieval effectiveness. In this study, we introduce new document priors and empirically investigate their effect by employing them as features in a learning to rank (LTR) deployment. The experiments are performed on the two standard Web IR test collections: the ClueWeb09 and the ClueWeb12 datasets, which include 500 and 733 million web documents, respectively, and the associated TREC & NTCIR query sets with a total number of 1,204 queries. A strong baseline is formed by using standard features introduced in the previous works, with respect to which the effect of newly introduced features in this paper is empirically compared. We test our features by LambdaMART, which is state-of-the-art LTR technique. The results reveal that the features introduced in this work led improvement in retrieval performance on the test collections in use. The introduced features are classified into 5 groups with respect to functional properties and each group is also analyzed in detail.

Locating Target Regions for Image Retrieval in an Unsupervised Manner.

Image retrieval performance can be improved by training a convolutional neural network (CNN) model with annotated data to facilitate accurate localization of target regions. However, obtaining sufficiently annotated data is expensive and impractical in real settings. It is challenging to achieve accurate localization of target regions in an unsupervised manner. To address this problem, we propose a new unsupervised image retrieval method named unsupervised target region localization (UTRL) descriptors. It can precisely locate target regions without supervisory information or learning. Our method contains three highlights: 1) we propose a novel zero-label transfer learning method to address the problem of co-localization in target regions. This enhances the potential localization ability of pretrained CNN models through a zero-label data-driven approach; 2) we propose a multiscale attention accumulation method to accurately extract distinguishable target features. It distinguishes the importance of features by using local Gaussian weights; and 3) we propose a simple yet effective method to reduce vector dimensionality, named twice-PCA-whitening (TPW), which reduces the performance degradation caused by feature compression. Notably, TPW is a robust and general method that can be widely applied to image retrieval tasks to improve retrieval performance. This work also facilitates the development of image retrieval based on short vector features. Extensive experiments on six popular benchmark datasets demonstrate that our method achieves about 7% greater mean average precision (mAP) compared to existing state-of-the-art unsupervised methods.

Incorporating Structural Information into Legal Case Retrieval

Legal case retrieval has received increasing attention in recent years. However, compared to ad hoc retrieval tasks, legal case retrieval has its unique challenges. First, case documents are rather lengthy and contain complex legal structures. Therefore, it is difficult for most existing dense retrieval models to encode an entire document and capture its inherent complex structure information. Most existing methods simply truncate part of the document content to meet the input length limit of PLMs, which will lead to information loss. Additionally, the definition of relevance in the legal domain differs from that in the general domain. Previous semantic-based or lexical-based methods fail to provide a comprehensive understanding of the relevance of legal cases. In this article, we propose a Structured Legal case Retrieval (SLR) framework, which incorporates internal and external structural information to address the above two challenges. Specifically, to avoid the truncation of long legal documents, the internal structural information, which is the organization pattern of legal documents, can be utilized to split a case document into segments. By dividing the document-level semantic matching task into segment-level subtasks, SLR can separately process segments using different methods based on the characteristic of each segment. In this way, the key elements of a case document can be highlighted without losing other content information. Second, toward a better understanding of relevance in the legal domain, we investigate the connections between criminal charges appearing in large-scale case corpus to generate a chargewise relation graph. Then, the similarity between criminal charges can be pre-computed as the external structural information to enhance the recognition of relevant cases. Finally, a learning-to-rank algorithm integrates the features collected from internal and external structures to output the final retrieval results. Experimental results on public legal case retrieval benchmarks demonstrate the superior effectiveness of SLR over existing state-of-the-art baselines, including traditional bag-of-words and neural-based methods. Furthermore, we conduct a case study to visualize how the proposed model focuses on key elements and improves retrieval performance.

Person-action Instance Search in Story Videos: An Experimental Study

Person-Action instance search (P-A INS) aims to retrieve the instances of a specific person doing a specific action, which appears in the 2019–2021 INS tasks of the world-famous TREC Video Retrieval Evaluation (TRECVID). Most of the top-ranking solutions can be summarized with a Division-Fusion-Optimization (DFO) framework, in which person and action recognition scores are obtained separately, then fused, and, optionally, further optimized to generate the final ranking. However, TRECVID only evaluates the final ranking results, ignoring the effects of intermediate steps and their implementation methods. We argue that conducting the fine-grained evaluations of intermediate steps of DFO framework will (1) provide a quantitative analysis of the different methods’ performance in intermediate steps; (2) find out better design choices that contribute to improving retrieval performance; and (3) inspire new ideas for future research from the limitation analysis of current techniques. Particularly, we propose an indirect evaluation method motivated by the leave-one-out strategy, which finds an optimal solution surpassing the champion teams in 2020–2021 INS tasks. Moreover, to validate the generalizability and robustness of the proposed solution under various scenarios, we specifically construct a new large-scale P-A INS dataset and conduct comparative experiments with both the leading NIST TRECVID INS solution and the state-of-the-art P-A INS method. Finally, we discuss the limitations of our evaluation work and suggest future research directions.

A DCT probability histogram-based ROI features for content-based natural and medical image retrieval applications

The effectiveness of a Content-based Image Retrieval (CBIR) system depends on selecting representative and salient visual features. Capturing the salient image features for forming a feature vector is an essential part of the CBIR process, as it can significantly affect retrieval efficiency. Most images contain a visually prominent region that perceives maximum image semantic information, known as Regions-of-Interest (ROIs). This paper suggests a new CBIR scheme that employs a Graph-based Visual Saliency (GBVS) map to extract the ROI of the image. Further, block-level Discrete Cosine Transforms (DCT) and subsequent histograms are computed for the DC and significant AC coefficients of the extracted ROI image. In addition, a low-dimensional feature vector is formed from the modified histograms. Subsequently, six statistical features from the background scene have also been computed for the final feature vector construction. As a result, the final feature vector contains ROI and background information in their proper proportion to improve retrieval performance. Moreover, the performance of the proposed CBIR scheme has been evaluated using one medical, three natural, two objects, one texture, and one natural heterogeneous image database. All these databases collectively contain 53732 different images from 200 image classes. The outcomes of the experiments demonstrate substantial improvement in the retrieval process with some existing related schemes. At the same time, it also shows that the performance of the proposed scheme is robust concerning all kinds of image databases.

SketchCADGAN: A generative approach for completing partially drawn query sketches of engineering shapes to enhance retrieval system performance

Retrieval systems are commonly used to find relevant data in large datasets. In engineering, these systems are useful for locating specific engineering shapes in a large dataset of engineering components. When end users want to search for a shape, they prefer a two-dimensional (2D) sketch over a three-dimensional (3D) object. However, users lacking domain knowledge may struggle to generate a complete query sketch and provide a partially completed sketch instead. Retrieving relevant information from partially drawn sketches is difficult because they may have missing edges, partially drawn circles, holes, ovals, etc. Most retrieval systems compare the similarity between the query and items in the database, so incomplete sketches may be ineffective in finding the relevant information.To address this problem with incomplete sketches, we propose a new generative adversarial network called SketchCADGAN. This network uses a two-stage cascaded architecture, with the first network attempting to predict a CAD model image from an incomplete sketch and the second network using the CAD model image to predict a completed sketch. Both networks are trained together adversarially. Our approach is proven more effective than other advanced techniques through qualitative and quantitative comparisons. Furthermore, we present the results of the retrieval system using both partially drawn and completed sketches, and demonstrate that incorporating completed sketches from the suggested cascaded GAN architecture results in improved retrieval performance.

BIM Search Engine: Effects of Object Relationships and Information Standards

As Building Information Modeling (BIM) models are getting bigger, with more information linked to geometrical 3D models, a dedicated BIM search engine is important. A BIM search engine was developed to examine the value of exploiting a 3D object’s topological relationships to other 3D objects when assessing that object’s relevance to a query. The impacts of two information standards, the Industry Foundation Classes (IFC) and Uniclass-2015, on information retrieval (IR) performance were also measured. The 3DIR Autodesk Revit toolset was used on a structural model of a 16-story building from an industry partner. The retrieval performance measures of Precision and Recall did not clearly highlight the benefit, although the increased relevance values of those objects deemed by experts to be relevant demonstrates the promise of such contextual measures. The effect of shifting from the native Revit file format to various standards was tested: IR performance was poor with the Uniclass-2015 dataset and was comparable to the Revit native model with the IFC model. Although not shown conclusively to improve retrieval performance, the contextual relevance measures presented in this paper are promising and constitute a significant original contribution. Future research is needed to fine-tune these measures and fully realize their potential.

Image-Text Retrieval With Cross-Modal Semantic Importance Consistency

Cross-modal image-text retrieval is an important area of Vision-and-Language task that models the similarity of image-text pairs by embedding features into a shared space for alignment. To bridge the heterogeneous gap between the two modalities, current approaches achieve inter-modal alignment and intra-modal semantic relationship modeling through complex weighted combinations between items. In the intra-modal association and inter-modal interaction processes, the higher-weight items have a higher contribution to the global semantics. However, the same item always produces different contributions in the two processes, since most traditional approaches only focus on the alignment. This usually results in semantic changes and misalignment. To address this issue, this paper proposes Cross-modal Semantic Importance Consistency (CSIC) which achieves invariance in the semantic of items during aligning. The proposed technique measures the semantic importance of items obtained from intra-modal and inter-modal self-attention and learns a more reasonable representation vector by inter-calibrating the importance distribution to improve performance. We conducted extensive experiments on the Flickr30K and MS COCO datasets. The results show that our approach can significantly improve retrieval performance, proving the proposed approach’s superiority and rationality.

Deep Learning of Image Representations with Convolutional Neural Networks Autoencoder for Image Retrieval with Relevance Feedback

mage retrieval with traditional relevance feedback encounters problems: (1) ability to represent handcrafted features which is limited, and (2) inefficient withhigh-dimensional data such as image data. In this paper,we propose a framework based on very deep convolutionalneural network autoencoder for image retrieval, called AIR(Autoencoders for Image Retrieval). Our proposed frameworkallows to learn feature vectors directly from the raw imageand in an unsupervised manner. In addition, our frameworkutilizes a hybrid approach of unsupervised and supervisedto improve retrieval performance. The experimental resultsshow that our method gives better results than some existingmethods on the CIFAR-100 image set, which consists of 60,000images.

Domain Adaptation Problem in Sketch Based Image Retrieval

In this article, we present two algorithms that discover the discriminative structures of sketches, given pairs of sketches and photos in sketch-based image retrieval (SBIR) scenarios. Unlike the existing approaches, we aim at the few-shot and domain adaptation (DA) problems, and set up a module with canonical correlation analysis (CCA) technique in our algorithms to improve retrieval performance. For single source domain settings, our first algorithm can effectively transfer a classifier trained on a known dataset to a new one. For multisource settings, our second algorithm sophisticatedly combines multisource domain data to yield a classifier on the target domain. To the best of our knowledge, these two works are the first research in SBIR field. Experiments on the Sketchy and TU-Berlin sketch benchmark datasets demonstrate the effectiveness of our algorithms and compelling performance. Compared with the state-of-the-art methods, our algorithms do not use text based semantic information, but achieve competitive results. Furthermore, experiments also turn out that feature representation by available trained deep networks has distinct advantages and combination with traditional machine learning methods brings substantial improvements against the state-of-the-art methods on image retrievals. The complete source code of the proposed algorithms will be released on gitHub: [ GitHub- ADAM0912/Domain-Adaptation-Problem-in-Sketch-Based-Image-Retrieval: The code repository for the articleDomain Adaptation Problem in Sketch Based Image Retrieval ] .

Hybrid Bag-of-Visual-Words and FeatureWiz Selection for Content-Based Visual Information Retrieval.

Recently, content-based image retrieval (CBIR) based on bag-of-visual-words (BoVW) model has been one of the most promising and increasingly active research areas. In this paper, we propose a new CBIR framework based on the visual words fusion of multiple feature descriptors to achieve an improved retrieval performance, where interest points are separately extracted from an image using features from accelerated segment test (FAST) and speeded-up robust features (SURF). The extracted keypoints are then fused together in a single keypoint feature vector and the improved RootSIFT algorithm is applied to describe the region surrounding each keypoint. Afterward, the FeatureWiz algorithm is employed to reduce features and select the best features for the BoVW learning model. To create the codebook, K-means clustering is applied to quantize visual features into a smaller set of visual words. Finally, the feature vectors extracted from the BoVW model are fed into a support vector machines (SVMs) classifier for image retrieval. An inverted index technique based on cosine distance metric is applied to sort the retrieved images to the similarity of the query image. Experiments on three benchmark datasets (Corel-1000, Caltech-10 and Oxford Flower-17) show that the presented CBIR technique can deliver comparable results to other state-of-the-art techniques, by achieving average accuracies of 92.94%, 98.40% and 84.94% on these datasets, respectively.

Robust Video-Text Retrieval Via Noisy Pair Calibration

Video-text retrieval is a fundamental task in managing the emerging massive amounts of video data. The main challenge focuses on learning a common representation space for videos and queries where the similarity measurement can reflect the semantic closeness. However, existing video-text retrieval models may suffer from the following noise in the common space learning procedure: First, the video-text correspondences in positive pairs may not be exact matches. The crowdsourcing annotation for existing datasets leads to inevitable tagging noise for non-expert annotators. Second, the learning of video-text representation is based on the negative samples randomly sampled. Instances that are semantically similar to the query may be incorrectly categorized as negative samples. To alleviate the adverse impact of these noisy pairs, we propose a novel robust video-text retrieval method that protects the model from noisy positive and negative pairs by identifying and calibrating noisy pairs with their uncertainty score. In particular, we propose a noisy pair identifier, which divides the training dataset into noisy and clean subsets based on the estimated uncertainty of each pair. Then, with the help of uncertainties, we calibrate the two types of noisy pairs with an adaptive margin triplet loss and a weighted triplet loss function, respectively. To verify the effectiveness of our methods, we conduct extensive experiments on three widely used datasets. Experimental results show that the proposed robust video-text retrieval methods successfully identify and calibrate the noisy pairs and improve retrieval performance.

DHAN: Encrypted JPEG image retrieval via DCT histograms-based attention networks

In image retrieval, the images to be retrieved are stored on remote servers. Since the images contain amounts of privacy information and the server cannot be fully trusted, people usually encrypt their images before uploading them to the server, which raises the demand for encrypted image retrieval (EIR). Current EIR techniques extract ruled hand-craft features from cipher images first and then build retrieval models (e.g., support vector machine, SVM) by these features, or use deep learning models (e.g., Convolutional Neural Network, CNN) to learn cipher-image representations in an end-to-end manner. However, SVM is not skilled at learning non-linear embedding in complex image database, and end-to-end EIR leads to low image security or retrieval performance because CNN is sensitive to extreme chaotic cipher images. Not-end-to-end EIR offers excellent encryption performance, and deep learning can further mine non-linear embedding from ruled hand-craft features. To this end, we propose a novel EIR scheme, named discrete cosine transform (DCT) Histograms-based Attention Networks (DHAN), which is based on deep learning to enhance expression ability of cipher-image in a not-end-to-end manner. Specifically, the DCT coefficients of images are encrypted by value replacement and block permutation encryption, and then the effective histogram features of DCT coefficients are extracted from the cipher images since the sets of DCT frequency in encrypted images are similar to that of plain images. After that, to dynamically learn the salient features of cipher images, we propose a new module named ResAttention and design deep attention networks to provide retrieval. Extensive experiments on two datasets demonstrate that DHAN not only provides high image security but also greatly improves retrieval performance than that of existing schemes.

Graph-based query reformulation system for descriptive queries of jargon words using definitions

In information retrieval, the lexical mismatch is a common problem where texts in documents and user queries have different forms even though they have the same meaning. The problem occurs more frequently when people search for jargon words because they tend to use descriptive queries and the jargon words interchangeably such as “hypertension” and “high blood pressure”. Query reformulation (QR) is one of the techniques that overcome the lexical gaps by transforming users’ descriptive queries into appropriate jargon queries. This study proposes a novel graph-based QR system that uses a dictionary, where the model does not require manually annotated data. Given a descriptive query, our system predicts the corresponding jargon word over a graph consisting of headwords and words in their definition. We use a graph neural network to represent the relational properties between words and the curriculum learning to train the networks stably and gradually adjust to hard samples. Moreover, we propose a graph search model that finds the target node in real-time using the relevance scores of neighborhood nodes. By adding this fast graph search model to the front of the proposed system, we reduce the reformulating time significantly. Experimental results on two datasets show that the proposed method can effectively reformulate descriptive queries to corresponding jargon words as well as improve retrieval performance under several search frameworks.

Deep metric learning with mirror attention and fine triplet loss for fundus image retrieval in ophthalmology

Fundus image retrieval can help ophthalmologists make evidence-based medico-decision by providing similar cases. Its basic task is to learn highly discriminative visual descriptors from image space, in which lesion features are the main differentiating clue. Lesions in fundus images appear small in size, similar in textures, and scatter around vessels, such as microaneurysms and hemorrhages. Hence, although a single small lesion has a saliently visual manifestation, its discriminative information is hard to reserve in the last image descriptors. For fundus images, the optic disc of the left and right eyes are symmetric, and the macular area lies in the central axis from the vertical view. Based on such spatial structure and lesion characteristics, we present a novel deep metric learning framework equipped with mirror attention to enhance the discriminative features of small and scattering lesions and encode them into image descriptors. The mirror attention can give lesions high attention scores by capturing spatial dependency of vertical and horizontal views, especially the relations between lesions and vessels. Based on the mirror attention, we further propose a new fine triplet loss to confine distances of positive pairs by exploiting the learned relevant degrees of positive pairs in a self-supervised manner. The fine triplet loss can help detect the subtle differences of positive pairs to improve the ranking performance of hit items. To demonstrate the effectiveness of improving retrieval performance, we conduct comprehensive experiments on the largest fundus dataset of diabetic retinopathy (DR) detection and achieve the best precision compared to counterparts. The experiments show that our method produces significant performance improvements for fundus image retrieval, especially the ranking quality of DR grades containing microaneurysms and hemorrhages. Our proposed mirror attention can be applied to off-the-shelf backbones and trained efficiently in an end-to-end manner for other medical images to obtain highly discriminative image descriptors.