Simple Concatenation Research Articles

Multi-attention guided feature fusion network for salient object detection

Though with the rapid development of deep learning, salient object detection methods have achieved increasingly better performance, how to get effective feature representation to predict more accurate saliency maps is still a burning problem we need to consider. To overcome this situation, most previous works tend to focus on skip-based architecture to integrate hierarchical information of different scales and layers. However, a simple concatenation of high-level features and low-level features is not all-powerful because cluttered and noisy information can cause negative consequences. Concerning the issue mentioned above, we propose a Multi-Attention guided Feature-fusion network (MAF) which can alleviate the problem from two aspects. For one thing, we use a novel Channel-wise Attention Block (CAB) to in charge of message passing layer by layer from a global view, which utilizes the semantic cues in the higher convolutional block to instruct the feature selection in the lower block. For another, a Position Attention Block (PAB) also works on integrated features to model pixel relationships and capture rich contextual dependencies. Under the guidance of multi-attention, discriminative features are selected to conduct a new end-to-end densely supervised encoder-decoder network which detects salient objects more uniformly and precisely. As the experimental results on five benchmark datasets show, our methods perform favorably against other state-of-the-art approaches.

Alternating Language Modeling for Cross-Lingual Pre-Training

Language model pre-training has achieved success in many natural language processing tasks. Existing methods for cross-lingual pre-training adopt Translation Language Model to predict masked words with the concatenation of the source sentence and its target equivalent. In this work, we introduce a novel cross-lingual pre-training method, called Alternating Language Modeling (ALM). It code-switches sentences of different languages rather than simple concatenation, hoping to capture the rich cross-lingual context of words and phrases. More specifically, we randomly substitute source phrases with target translations to create code-switched sentences. Then, we use these code-switched data to train ALM model to learn to predict words of different languages. We evaluate our pre-training ALM on the downstream tasks of machine translation and cross-lingual classification. Experiments show that ALM can outperform the previous pre-training methods on three benchmarks.1

Multi-layer manifold learning with feature selection

Many fundamental problems in machine learning require some form of dimensionality reduction. To this end, two different strategies were used: Manifold Learning and Feature Selection. Manifold learning (or data embedding) attempts to compute a subspace from original data by feature recombination/transformation. Feature selection aims to select the most relevant features in the original space. In this paper, we propose a novel cooperative Manifold learning-Feature selection that goes beyond the simple concatenation of these two modules. Our basic idea is to transform a given shallow embedding to a deep variant by computing a cascade of embeddings in which each embedding undergoes feature selection and elimination. We use filter approaches in order to efficiently select irrelevant features at any stage of the process. For a case study, our proposed framework was used with two typical linear embedding algorithms: Local Discriminant Embedding (LDE) (a supervised technique) and Locality Preserving Projections (LPP) (unsupervised technique) on four challenging face databases and it has been conveniently compared with other cooperative schemes. Moreover, a comparison with several state-of-the-art manifold learning methods is provided. As it is exhibited by our experimental study, the proposed framework can achieve superior learning performance with respect to classic cooperative schemes and to many competing manifold learning methods.

Multi-level supervised hashing with deep features for efficient image retrieval

Image hashing based on deep convolutional neural networks (CNN), deep hashing, has acquired breakthrough in image retrieval. Although deep features from various CNN layers have various levels of information, most of the existing deep hashing methods extract the feature vector only from the output of the penultimate fully-connected layer, focusing primarily on semantic information whilst ignoring detailed structure information. This calls for research on multi-level hashing, utilizing multi-level features to exploit different levels of CNN characteristics. To fill this gap, a novel image hashing method, Multi-Level Supervised Hashing with deep feature (MLSH), is proposed in this paper to further exploit multiple levels of deep image features. It uses a multiple-hash-table mechanism to integrate multi-level features extracted from an individual deep convolutional neural network. It takes advantage of the complementarity among multi-level features from various layers of a single deep network. High-level features reveal the semantic content of the image, while low-level features provide the structural information that is missing in high-level features. Instead of simple concatenation, several hash tables are trained individually using different levels of features from different layers, which are then integrated for efficient image retrieval. The method has been systematically evaluated through experiments on three image databases, including CIFAR-10, MNIST and NUSWIDE, and has thus been demonstrated to set a new state of the art in image hashing, outperforming several state-of-the-art hashing methods. Furthermore, the recall and precision can be balanced and improved simultaneously.

Classification of stomach infections: A paradigm of convolutional neural network along with classical features fusion and selection.

Automated detection and classification of gastric infections (i.e., ulcer, polyp, esophagitis, and bleeding) through wireless capsule endoscopy (WCE) is still a key challenge. Doctors can identify these endoscopic diseases by using the computer-aided diagnostic (CAD) systems. In this article, a new fully automated system is proposed for the recognition of gastric infections through multi-type features extraction, fusion, and robust features selection. Five key steps are performed-database creation, handcrafted and convolutional neural network (CNN) deep features extraction, a fusion of extracted features, selection of best features using a genetic algorithm (GA), and recognition. In the features extraction step, discrete cosine transform, discrete wavelet transform strong color feature, and VGG16-based CNN features are extracted. Later, these features are fused by simple array concatenation and GA is performed through which best features are selected based on K-Nearest Neighbor fitness function. In the last, best selected features are provided to Ensemble classifier for recognition of gastric diseases. A database is prepared using four datasets-Kvasir, CVC-ClinicDB, Private, and ETIS-LaribPolypDB with four types of gastric infections such as ulcer, polyp, esophagitis, and bleeding. Using this database, proposed technique performs better as compared to existing methods and achieves an accuracy of 96.5%.

Microscopic melanoma detection and classification: A framework of pixel-based fusion and multilevel features reduction.

The numbers of diagnosed patients by melanoma are drastic and contribute more deaths annually among young peoples. An approximately 192,310 new cases of skin cancer are diagnosed in 2019, which shows the importance of automated systems for the diagnosis process. Accordingly, this article presents an automated method for skin lesions detection and recognition using pixel-based seed segmented images fusion and multilevel features reduction. The proposed method involves four key steps: (a) mean-based function is implemented and fed input to top-hat and bottom-hat filters which later fused for contrast stretching, (b) seed region growing and graph-cut method-based lesion segmentation and fused both segmented lesions through pixel-based fusion, (c) multilevel features such as histogram oriented gradient (HOG), speeded up robust features (SURF), and color are extracted and simple concatenation is performed, and (d) finally variance precise entropy-based features reduction and classification through SVM via cubic kernel function. Two different experiments are performed for the evaluation of this method. The segmentation performance is evaluated on PH2, ISBI2016, and ISIC2017 with an accuracy of 95.86, 94.79, and 94.92%, respectively. The classification performance is evaluated on PH2 and ISBI2016 dataset with an accuracy of 98.20 and 95.42%, respectively. The results of the proposed automated systems are outstanding as compared to the current techniques reported in state of art, which demonstrate the validity of the proposed method.

Intelligent Fusion of Deep Features for Improved Waste Classification

In this article, we address the problem of an image-based automatic classification of waste materials. Given the large number of waste categories and the importance of proper management of waste materials, the problem is known to be critical and of a particular interest. To achieve reliable waste classification capability, we propose a novel approach, that we name double fusion, which optimally combines multiple deep learning models using feature and score-level fusion methods. The double fusion scheme ensures an optimized contribution of the deep models by, firstly, combining their capabilities in an early and late fusion scheme followed by a score-level fusion of the classification results obtained with early and late fusion methods. In total, we employ and compare six different fusion methods including two feature-level fusion schemes, namely (i) Discriminant Correlation Analysis and (ii) simple concatenation of deep features, and four late fusion methods, namely (i) Particle Swarm Optimization, (ii) Genetic modeling of deep features (iii) Induced Ordered Weighted Averaging and (iv) a baseline method where all the deep models are treated equally. Moreover, we also evaluate the performance of the individual deep models, and compare our results against state-of-the-art methods demonstrating a significant improvement of 3.58% over state-of-the-art.

Cancelable Face Template Protection using Transform Features for Cyberworld Security

Cyber world becomes a fundamental and vital component of the physical world with the increase of dependence on internet-connected devices in industry and government organizations. Provision of privacy and security of users during online communication offers unique cybersecurity challenges for industry and government. Intrusion is one of the crucial issues of cybersecurity, which can be overcome by providing the vigorous authentication solutions. Biometrics authentication is used in different cybersecurity systems for user authentication purpose. The cancelable biometric is a solution to rid of privacy problems in traditional biometric systems. This paper purposes a new cancelable face authentication method, which uses Hybrid Gabor PCA (HGPCA) descriptor for cyberworld security. The proposed method uses the wavelet transform for the extraction of the features of the face images by using Gabor filter and Principal Component Analysis (PCA). Later, both types of features have been ensemble by using the simple concatenation scheme. Then scrambling has been applied to the fused features by using the random key generated by the user. So finally, scrambled fused features has been stored in the database which are used for the cancelable biometric authentication as well as recovery. HGPCA achieves “cancelability” and increases the authentication accuracy. The proposed method has been tested on three standard face datasets. Experimental results of the proposed method have been compared with existing methods by using standard quantitative measures that show superiority over existing methods.

Part-of-Speech Tagging Using Multiview Learning

In natural language processing, character-level representations are vector representations of the particular character. Character-level representations have recently focused on enriching subword information by stacking deep neural models. Ideally, applications of several character-level representations can help capture different aspects of the subword information. However, this approach has often failed in the past, mainly because of the nature of traditionally used simple concatenation models. In this study, we explore different character-level modeling techniques. During the learning process, long short-term memory-based character representations can introduce different views for a part-of-speech tagger. After investigating two previously reported techniques, we propose two additional extended methods: (1) a multihead-attention character-level representation for capturing several aspects of subword information, and (2) an optimal structure for training two different character-level embeddings based on joint learning. We evaluate our results on the part-of-speech (POS) tagging dataset of the Conference on Natural Language Learning (CoNLL) 2018 shared task in universal dependencies. We show that our method substantially improves POS tagging results for many morphologically rich languages where the character information should be considered more substantially. Moreover, we compare the performance of our model with recent state-of-the-art POS taggers, which are trained with language models such as Bidirectional Encoder Representations from Transformers (BERT) and Deep Contextualized Word Representations (ELMo); our multiview tagger shows better results for nine languages. The proposed character model shows significant improvements in Ancient Greek, with average gains of 8.89 points in accuracy compared to the previous word representation model. Therefore, our empirical experiments indicate that character-level representations are more important than word representations for morphologically rich languages in terms of performance.

Categorical Metadata Representation for Customized Text Classification

The performance of text classification has improved tremendously using intelligently engineered neural-based models, especially those injecting categorical metadata as additional information, e.g., using user/product information for sentiment classification. This information has been used to modify parts of the model (e.g., word embeddings, attention mechanisms) such that results can be customized according to the metadata. We observe that current representation methods for categorical metadata, which are devised for human consumption, are not as effective as claimed in popular classification methods, outperformed even by simple concatenation of categorical features in the final layer of the sentence encoder. We conjecture that categorical features are harder to represent for machine use, as available context only indirectly describes the category, and even such context is often scarce (for tail category). To this end, we propose using basis vectors to effectively incorporate categorical metadata on various parts of a neural-based model. This additionally decreases the number of parameters dramatically, especially when the number of categorical features is large. Extensive experiments on various data sets with different properties are performed and show that through our method, we can represent categorical metadata more effectively to customize parts of the model, including unexplored ones, and increase the performance of the model greatly.

GF-Net: Improving machine reading comprehension with feature gates

Machine reading comprehension (MRC) is a field of question-answering in which computers understand given passages and answer related questions. Several previous models have tried to combine the use of linguistic and word embedding features to improve the performance of MRC; however, they could not obtain successful results because of feature interference problems caused by simple concatenation of the two. To resolve these problems, a machine reading comprehension model called gated feature network (GF-Net) is proposed in which linguistic features are selectively used according to their roles in the process of answer selection. In the GF-Net, the weights of the linguistic features are automatically controlled through gate mechanisms called feature gates. In the experiments with Stanford Question Answering Dataset SQuAD, the MRC models with feature gates showed a 0.67%p higher average of exact match (EM) and 0.64%p higher average of F1-score than models without feature gates. In addition, the GF-Net outperformed the previous MRC models to which feature gates were added. Based on these experimental results, it is concluded that the gate mechanism can contribute to an improvement in the performance of MRC models and the architecture of the GF-Net is suitable for the task of MRC.

Multimodal biometric system combining left and right palmprints

PurposePalmprint recognition is a very interesting and promising area of research. Much work has already been done in this area, but much more needs to be done to make the systems more efficient. In this paper, a multimodal biometrics system based on fusion of left and right palmprints of a person is proposed to overcome limitations of unimodal systems.Design/methodology/approachFeatures are extracted using some proposed multi-block local descriptors in addition to MBLBP. Fusion of extracted features is done at feature level by a simple concatenation of feature vectors. Then, feature selection is performed on the resulting global feature vector using evolutionary algorithms such as genetic algorithms and backtracking search algorithm for a comparison purpose. The benefits of such step selecting the relevant features are known in the literature, such as increasing the recognition accuracy and reducing the feature set size, which results in runtime saving. In matching step, Chi-square similarity measure is used.FindingsThe resulting feature vector length representing a person is compact and the runtime is reduced.Originality/valueIntensive experiments were done on the publicly available IITD database. Experimental results show a recognition accuracy of 99.17 which prove the effectiveness and robustness of the proposed multimodal biometrics system than other unimodal and multimodal biometrics systems.

Improving plankton image classification using context metadata

AbstractAdvances in both hardware and software are enabling rapid proliferation of in situ plankton imaging methods, requiring more effective machine learning approaches to image classification. Deep Learning methods, such as convolutional neural networks (CNNs), show marked improvement over traditional feature‐based supervised machine learning algorithms, but require careful optimization of hyperparameters and adequate training sets. Here, we document some best practices in applying CNNs to zooplankton and marine snow images and note where our results differ from contemporary Deep Learning findings in other domains. We boost the performance of CNN classifiers by incorporating metadata of different types and illustrate how to assimilate metadata beyond simple concatenation. We utilize both geotemporal (e.g., sample depth, location, time of day) and hydrographic (e.g., temperature, salinity, chlorophyll a) metadata and show that either type by itself, or both combined, can substantially reduce error rates. Incorporation of context metadata also boosts performance of the feature‐based classifiers we evaluated: Random Forest, Extremely Randomized Trees, Gradient Boosted Classifier, Support Vector Machines, and Multilayer Perceptron. For our assessments, we use an original data set of 350,000 in situ images (roughly 50% marine snow and 50% non‐snow sorted into 26 categories) from a novel in situ Zooglider. We document asymptotically increasing performance with more computationally intensive techniques, such as substantially deeper networks and artificially augmented data sets. Our best model achieves 92.3% accuracy with our 27‐class data set. We provide guidance for further refinements that are likely to provide additional gains in classifier accuracy.

A deep one-shot network for query-based logo retrieval

Logo detection in real-world scene images is an important problem with applications in advertisement and marketing. Existing general-purpose object detection methods require large training data with annotations for every logo class. These methods do not satisfy the incremental demand of logo classes necessary for practical deployment since it is practically impossible to have such annotated data for new unseen logo. In this work, we develop an easy-to-implement query-based logo detection and localization system by employing a one-shot learning technique using off the shelf neural network components. Given an image of a query logo, our model searches for logo within a given target image and predicts the possible location of the logo by estimating a binary segmentation mask. The proposed model consists of a conditional branch and a segmentation branch. The former gives a conditional latent representation of the given query logo which is combined with feature maps of the segmentation branch at multiple scales in order to obtain the matching location of the query logo in a target image. Feature matching between the latent query representation and multi-scale feature maps of segmentation branch using simple concatenation operation followed by 1 × 1 convolution layer makes our model scale-invariant. Despite its simplicity, our query-based logo retrieval framework achieved superior performance in FlickrLogos-32 and TopLogos-10 dataset over different existing baseline methods.

Pedestrian detection using first- and second-order aggregate channel features

The content-based analysis of visual multimedia like images and videos are urgently needed to empower human society for the automation of difficult tasks. Pedestrian detection serves as a backbone for a multitude of image processing and machine learning algorithms and secures quite a lot of real-world applications. Keeping this fact in mind, here, we deal with the fabrication of suitable features to identify human/pedestrian instances from images with near accuracy. Accordingly, we introduce second-order aggregate channel features (SOACF) to enhance the performance of much-celebrated pedestrian detection algorithm which was mainly based on the first-order information in an image—aggregate channel features detector (ACF detector). We experimentally proved the complementary nature of ACF and SOACF. Designed to garner both these features together, instead of simple concatenation, or direct merging of the two detectors, we employed a weighted non-maximum suppression merging algorithm. The prospective detector not only performed well on INRIA, Caltech and KITTI pedestrian data set but also, mitigate the miss rate by $$\sim 4\%$$ in Caltech data set and $$\sim 2\%$$ in KITTI data set in comparison with ACF detector. Despite the fact that our in-house generated detector uses only a few channels, it surpasses many state-of-the-art methods based on baseline ACF detector. Moreover, the detection speed is 100 times faster than the topmost pedestrian detector based on ACF.

On the Entropy Loss and Gap of Condensers

Many algorithms are proven to work under the assumption that they have access to a source of random, uniformly distributed bits. However, in practice, sources of randomness are often imperfect, giving n random bits that have only k < n min-entropy. The value n − k is called the entropy gap of the source. Randomness condensers are hash functions that hash any such source to a shorter source with reduced entropy gap g . The goal is to lose as little entropy as possible in this process. Condensers also have an error parameter ε and use a small seed of uniformly distributed bits whose length we desire to minimize as well. In this work, we study the exact dependencies between the different parameters of seeded randomness condensers. We obtain a non-explicit upper bound, showing the existence of condensers with entropy loss log (1+log 1/ε / g ) + O (1) and seed length log ( n − k / ε g ) + O (1). In particular, this implies the existence of condensers with O (log 1 / ε) entropy gap and constant entropy loss. This extends (with slightly improved parameters) the non-explicit upper bound for condensers presented in the work of Dodis et al. (2014), which gives condensers with entropy loss at least log log 1 / ε. We also give a non-explicit upper bound for lossless condensers, which have entropy gap g ≥ log 1 / ε / ε + O (1) and seed length log ( n − k / ε 2 g ) + O (1). Furthermore, we address an open question raised in (Dodis et al. 2014), where Dodis et al. showed an explicit construction of condensers with constant gap and O (log log 1/ ε) loss, using seed length O ( n log 1 / ε). In the same article they improve the seed length to O ( k log k ) and ask whether it can be further improved. In this work, we reduce the seed length of their construction to O (log ( n / ε)log ( k / &epsiv)) by a simple concatenation. In the analysis, we use and prove a tight equivalence between condensers and extractors with multiplicative error. We note that a similar, but non-tight, equivalence was already proven by Dodis et al. (Dodis et al. 2014) using a weaker variant of extractors called unpredictability extractors . We also remark that this equivalence underlies the work of Ben-Aroya et al. (Ben-Aroya et al. 2016) and later work on explicit two-source extractors, and we believe it is interesting in its own right.

A constraint-based programming approach for robotic assembly skills implementation

• A constraint-based programming formalism for the task-level implementation of robotic assembly skills that allows the skill developer to integrate force control requirements within the skill specification, independently of the specific robotic platform and assembly application. • Reactive force controlled motions are generated in real-time based on force-related objectives and constraints. • The proposed framework opens up the possibility for a non-expert factory worker to intuitively program complex assembly task by simple concatenation of assembly skills. • Demonstrations of a complete assembly process performed with the ABB YuMi dualarm robot in a sensorless configuration. The features of modern collaborative robots, mainly their kinematic redundancy combined with the light-weight structure, can be fully exploited in parts assembly. Traditional robot-level paradigm to robot programming, that requires to explicitly specify the motion of the robot and allows to use contact forces for motion supervision only, cannot be easily applied to complex interaction tasks, such as robotic assembly. Instead, by shifting paradigm to skill-based programming, it is possible to specify force control actions at task level and inherently provide compliant capabilities, without the need to specify the motions of the robot. To this end, this paper presents a constraint-based programming framework for the implementation of assembly skills for light-weight redundant robots, enabling a reactive generation of motion trajectories based on force control requirements. The effectiveness of the proposed approach is experimentally validated on a bimanual assembly use case performed with the ABB YuMi dual-arm robot, requiring a peg-in-hole insertion and a cap-rotation task. Estimation of the interaction force/torque additionally enables the execution of the assembly operation without the need for exteroceptive sensors.

Prediction of apoptosis protein subcellular localization via heterogeneous features and hierarchical extreme learning machine

ABSTRACTApoptosis is a fundamental process controlling normal tissue homeostasis by regulating a balance between cell proliferation and death. Predicting the subcellular location of apoptosis proteins is very helpful for understanding the mechanism of programmed cell death. Predicting protein subcellular localization with bioinformatic techniques provides quite a few opportunities in related fields. In this work, we propose the use of a hierarchical extreme learning machine (H-ELM) to make a classification of high-dimensional input data without demanding a dimension reduction process, which yields acceptable results. An attempt is made to extract features from different perspectives, and a feature fusion process is accomplished. Regarding the position-specific scoring matrix, the first type depicts the correlation within the sequence with the autocorrelation function for relatively random sections from the sequence; and the second type is the Kullback-Leibler (K-L) divergence of the two distributions formed by the amino acids’ constitutuent proportions. It is illustrated in an experiment with features from different sources mixed by simple concatenation yielding a poor result, but the synthetical feature fused with stochastic nonlinear embedding (t-SNE) greatly improved the classification. Finally, the highest overall accuracy of ZD98 is 87.5% by adjusting the hyper-parameters of H-ELM, and of CL317 is 92.4%.

Multisource Hyperspectral and LiDAR Data Fusion for Urban Land-Use Mapping based on a Modified Two-Branch Convolutional Neural Network

Accurate urban land-use mapping is a challenging task in the remote-sensing field. With the availability of diverse remote sensors, synthetic use and integration of multisource data provides an opportunity for improving urban land-use classification accuracy. Neural networks for Deep Learning have achieved very promising results in computer-vision tasks, such as image classification and object detection. However, the problem of designing an effective deep-learning model for the fusion of multisource remote-sensing data still remains. To tackle this issue, this paper proposes a modified two-branch convolutional neural network for the adaptive fusion of hyperspectral imagery (HSI) and Light Detection and Ranging (LiDAR) data. Specifically, the proposed model consists of a HSI branch and a LiDAR branch, sharing the same network structure to reduce the time cost of network design. A residual block is utilized in each branch to extract hierarchical, parallel, and multiscale features. An adaptive-feature fusion module is proposed to integrate HSI and LiDAR features in a more reasonable and natural way (based on “Squeeze-and-Excitation Networks”). Experiments indicate that the proposed two-branch network shows good performance, with an overall accuracy of almost 92%. Compared with single-source data, the introduction of multisource data improves accuracy by at least 8%. The adaptive fusion model can also increase classification accuracy by more than 3% when compared with the feature-stacking method (simple concatenation). The results demonstrate that the proposed network can effectively extract and fuse features for a better urban land-use mapping accuracy.

An out-of-band mobile authenticating mechanism for controlling access to data outsourced in the mobile cloud environment

'Mobile Cloud Computing' (MCC), a cloud environment, formed by mobile users at the client-side and cloud servers at the back-end enables users to store and pervasively access a huge amount of data via different mobile devices in a distributive manner. This increases the risk of data abuse, since data is outsourced on the distrusted cloud servers. The drawbacks of the current frameworks, such as overloaded computations in key distribution, reduced flexibility, and scalability, are unable to achieve fine-graininess and confidentiality. Thus, we propose an 'access control mechanism' which is lightweight with minimal computational overhead and provides fine-grained access control for sharing data using out-of-band (OOB) mobile authentication. Here we perform client-side encryption and decryption using simple hash functions and concatenation operator and achieve dynamic scalability. The encryption algorithms, sharing mechanism, and the use of OOB mobile authentication are extensively analysed to prove its efficiency and applicability.