Two-stage Classification Research Articles

A Cascaded Ensemble of Sparse-and-Dense Dictionaries for Vehicle Detection

Vehicle detection as a special case of object detection has practical meaning but faces challenges, such as the difficulty of detecting vehicles of various orientations, the serious influence from occlusion, the clutter of background, etc. In addition, existing effective approaches, like deep-learning-based ones, demand a large amount of training time and data, which causes trouble for their application. In this work, we propose a dictionary-learning-based vehicle detection approach which explicitly addresses these problems. Specifically, an ensemble of sparse-and-dense dictionaries (ESDD) are learned through supervised low-rank decomposition; each pair of sparse-and-dense dictionaries (SDD) in the ensemble is trained to represent either a subcategory of vehicle (corresponding to certain orientation range or occlusion level) or a subcategory of background (corresponding to a cluster of background patterns) and only gives good reconstructions to samples of the corresponding subcategory, making the ESDD capable of classifying vehicles from background even though they exhibit various appearances. We further organize ESDD into a two-level cascade (CESDD) to perform coarse-to-fine two-stage classification for better performance and computation reduction. The CESDD is then coupled with a downstream AdaBoost process to generate robust classifications. The proposed CESDD model is used as a window classifier in a sliding-window scan process over image pyramids to produce multi-scale detections, and an adapted mean-shift-like non-maximum suppression process is adopted to remove duplicate detections. Our CESDD vehicle detection approach is evaluated on KITTI dataset and compared with other strong counterparts; the experimental results exhibit the effectiveness of CESDD-based classification and detection, and the training of CESDD only demands small amount of time and data.

Outlier detection in test samples and supervised training set selection

‎Outlier detection is a technique for recognizing samples out of the main population within a data set‎. ‎Outliers have negative impacts on classification‎. ‎The recognized outliers are deleted to improve the classification power generally‎. ‎This paper proposes a method for outlier detection in test samples besides a supervised training set selection‎. ‎Training set selection is done based on the intersection of three well known similarity measures namely‎, ‎jacquard‎, ‎cosine‎, ‎and dice‎. ‎Each test sample is evaluated against the selected training set for possible outlier detection‎. ‎The selected training set is used for a two-stage classification‎. ‎The accuracy of classifiers are increased after outlier deletion‎. ‎The majority voting function is used for further improvement of classifiers‎.

Comparison of two object-oriented technologies for detecting land use change

This study compared two object-oriented land use change detection methods, namely detection after classification (DAC) and classification after detection (CAD) using digital elevation model, slope data, multi-temporal Thematic Mapper (TM) images in 2000 and Enhanced TM Plus images (ETM+) in 2010. Results revealed that the accuracy of the former was higher than that of the latter. A minimal discrepancy was observed in the accuracies of the two methods in detecting deciduous broad-leaved, evergreen coniferous and mixed forests; uplands; paddies; bare lands and human habitation, which can be extracted using spectral indices due to large spectrum differences. The accuracy of DAC was higher than that of CAD for areas with similar spectra, such as industrial and traffic lands, green shrubs, rivers, channels, reservoirs/pools and lake, because the former can fully utilise various information, including spectral, spatial, structural, textural and contextual information, during the two-stage classification. Moreover, the change area boundary was not limited and adjustable during the classification process. DAC can overcome the smoothing effect to a great extent using multi-scale segmentations and multiple indicators for detection. Although DAC can yield better results than CAD can, the former was more time-consuming than the latter because of its two-stage classification. Therefore, the hybrid of the two methods could be the appropriate approach.

The two-stage classification based on 1-SVM and RF classifiers

The approach to the two-stage classification based on the 1-SVM classifier used as the main classifier and the RF classifier used as the auxiliary classifier has been considered. The proposed approach improves the quality of classification when using imbalanced datasets. The results of the comparative analysis of the proposed approach and the alternative approach to the two-stage classification, in which the binary SVM classifier is used as the main classifier, and the RF classifier as the auxiliary classifier, are presented.

Two-stage classifier for face recognition using HOG features

Facerecognition ismainlyused in biometric applications, surveillance systems and computer vision.Themajorproblem while recognizingfacesarisesdue topose variations, background illumination invariants and facial expressions.Theproposed two-stage classifier aims to recognize given inputfaceimages usingk-NNand SVM. The use of two separate classifiers, one after the other, improves the system recognition accuracy.The pre processed face images are then utilized toextractprominent features by Histogram of Oriented Gradients (HOG).In the initial stage, k-NN classifier was used and afterwards, unrecognized face images were tested with SVM classifier.It is found that recognition accuracy increased by cascading k-NN followed bySVM.The two-stage classifierachieved a recognition accuracy of 95.2%.

On the Generalization and Reliability of Single Radar-Based Human Activity Recognition

Identifying human activities using short-range and low-power radars has attracted much attention among the researchers and consumer electronics industry. This paper considers human activity recognition in the context of a single Frequency Modulated Continuous Wave (FMCW) radar as the measurement tool. A classification pipeline is proposed to handle the data pre-processing and feature extraction and a machine-learning based solution is devised to undertake the activity classification. The performance of the proposed architecture is evaluated under both unseen subjects and new room layouts. We show how the accuracy of the activity classification will be affected by situations such as poor aspect-angle and occlusions created by furniture that normally arise in realistic scenarios where an unseen layout is considered. A two-stage classifier will be then proposed to enhance the generalization of the model, especially, to unseen rooms. Besides, an extensive feature exploration will be conducted and the importance of features in the generalization will be studied. The results in this paper will conclude a machine learning pipeline that will generalize well to unseen subjects and new room layouts, which are two main difficulties that arise in most radar-based activity classification tasks.

Investigating the use of a two-stage attention-aware convolutional neural network for the automated diagnosis of otitis media from tympanic membrane images: a prediction model development and validation study

ObjectivesThis study investigated the usefulness and performance of a two-stage attention-aware convolutional neural network (CNN) for the automated diagnosis of otitis media from tympanic membrane (TM) images.DesignA classification model development...

A Hybrid Two-Stage GNG–Modified VGG Method for Bone X-Rays Classification and Abnormality Detection

This paper introduces a novel, reliable, hybrid, two-stage method for bone x-ray classification and abnormality detection. The proposed method considers bones from seven areas of the upper extremities: namely, shoulder, humerus, forearm, elbow, wrist, hand, and finger. The novelty of the proposed work is not only in the application, but also in the proposed method itself by combining GNG network and eight models built from scratch and inspired from VGG model to achieve best performance and least computations possible. The features extracted from GNG are fed into a two-stage classification step. The first stage classifies a bone X-ray into one of seven types, after which it is directed according to type to one of seven classifiers trained to detect bone abnormality. Hence, the classification step consists of eight different models: one for classification and seven for abnormality detection. Experiments have been carried out using the MURA database, the largest public dataset of bone x-ray images. The best average sensitivity and specificity obtained for the first stage is 95.86% and 99.63%, respectively. For the second stage, the best average sensitivity and specificity obtained is 92.50% and 92.12%, respectively. These results are superior compared to state of art pretrained models. In addition, the computation and processing time are significantly decreased by the proposed scheme. Furthermore, to the best knowledge of authors, the proposed method is the first to integrate seven bones together in the same scheme. Finally, the hierarchical nature of the proposed method allows considering two problems together: bone classification and abnormality detection.

Two-Branch Relational Prototypical Network for Weakly Supervised Temporal Action Localization.

As a challenging task of high-level video understanding, weakly supervised temporal action localization has attracted more attention recently. With only video-level category labels, this task should indistinguishably identify the background and action categories frame by frame. However, it is non-trivial to achieve this in untrimmed videos, due to the unconstrained background, complex and multi-label actions. With the observation that these difficulties are mainly brought by the large variations within background and actions, we propose to address these challenges from the perspective of modeling variations. Moreover, it is desired to further reduce the variations, or learn compact features, so as to cast the problem of background identification as rejecting background and alleviate the contradiction between classification and detection. Accordingly, in this paper, we propose a two-branch relational prototypical network. The first branch, namely action-branch, adopts class-wise prototypes and mainly acts as an auxiliary to introduce priori knowledge about label dependencies and be a guide for the second branch. Meanwhile, the second branch, namely sub-branch, starts with multiple prototypes, namely sub-prototypes, to enable a powerful ability of modeling variations. As a further benefit, we elaborately design a multi-label clustering loss based on the sub-prototypes to learn compact features under the multi-label setting. The two branches are associated using the correspondences between two types of prototypes, leading to a special two-stage classifier in the s-branch, on the other hand, the two branches serve as regularization terms to each other, improving the final performance. Ablation studies find that the proposed model is capable of modeling classes with large variations and learning compact features. Extensive experimental evaluations on Thumos14, MultiThumos and ActivityNet datasets demonstrate the effectiveness of the proposed method and superior performance over state-of-the-art approaches.

Two-Stage Classification Method for MSI Status Prediction Based on Deep Learning Approach

Colorectal cancer is one of the most common cancers with a high mortality rate. The determination of microsatellite instability (MSI) status in resected cancer tissue is vital because it helps diagnose the related disease and determine the relevant treatment. This paper presents a two-stage classification method for predicting the MSI status based on a deep learning approach. The proposed pipeline includes the serial connection of the segmentation network and the classification network. In the first stage, the tumor area is segmented from the given pathological image using the Feature Pyramid Network (FPN). In the second stage, the segmented tumor is classified as MSI-L or MSI-H using Inception-Resnet-V2. We examined the performance of the proposed method using pathological images with 10× and 20× magnifications, in comparison with that of the conventional multiclass classification method where the tissue type is identified in one stage. The F1-score of the proposed method was higher than that of the conventional method at both 10× and 20× magnifications. Furthermore, we verified that the F1-score for 20× magnification was better than that for 10× magnification.

Angular Deep Supervised Vector Quantization for Image Retrieval.

Most of the deep quantization methods adopt unsupervised approaches, and the quantization process usually occurs in the Euclidean space on top of the deep feature and its approximate value. When this approach is applied to the retrieval tasks, since the internal product space of the retrieval process is different from the Euclidean space of quantization, minimizing the quantization error (QE) does not necessarily lead to a good performance on the maximum inner product search (MIPS). To solve these problems, we treat Softmax classification as vector quantization (VQ) with angular decision boundaries and propose angular deep supervised VQ (ADSVQ) for image retrieval. Our approach can simultaneously learn the discriminative feature representation and the updatable codebook, both lying on a hypersphere. To reduce the QE between centroids and deep features, two regularization terms are proposed as supervision signals to encourage the intra-class compactness and inter-class balance, respectively. ADSVQ explicitly reformulates the asymmetric distance computation in MIPS to transform the image retrieval process into a two-stage classification process. Moreover, we discuss the extension of multiple-label cases from the perspective of quantization with binary classification. Extensive experiments demonstrate that the proposed ADSVQ has excellent performance on four well-known image data sets when compared with the state-of-the-art hashing methods.

Gesture-Radar: A Dual Doppler Radar Based System for Robust Recognition and Quantitative Profiling of Human Gestures

Gesture recognition is key to enabling natural human-computer interactions. Existing approaches based on wireless sensing focus on accurate identification of arm gesture types. It remains a challenge to recognize and profile the details of arm gestures for precise interaction control. In addition, current approaches have strict positioning requirements between radars and users, making them difficult for real-world deployment. In this article, we adopt the multisensor approach and present gesture-radar-a dual Doppler radar-based gesture recognition and profiling system, which can capture subtle arm gestures with less positioning or environmental dependence. Gesture-radar uses two vertically placed Doppler radars to collect complementary sensing data of gestures, based on which cross-analysis can be performed for gesture recognition and profiling. Specifically, we first propose a two-stage classification model and enhance the signal proximity matching method by applying constraint functions to the DTW algorithm, aiming to improve the accuracy of gesture type recognition. Afterward, we establish and analyze unique features from the time-frequency spectrogram, which can be used to characterize in-depth gesture details, e.g., the angle or range of an arm movement. Experimental results show that gesture-radar achieves up to 93.5% average accuracy for gesture type recognition, and over 80% precision for profiling gesture details. This proves that the proposed approach is viable and can work in real-world environments.

Hybrid Two-Stage Architecture for Tampering Detection of Chipless ID Cards

Identity verification systems are widely used in daily life. Most of these systems rely on official documents containing identifying information about a person (i.e., passports, ID cards, driving licenses, amongst others). In this kind of approach, the identifiable data is contained inside the embedded chip in the ID card, and can be read remotely by an NFC-enabled mobile device and then matched with a frontal face photograph (selfie) of the person in question. Unfortunately, this method is limited in South-American countries, since only a few of them provide national ID cards that include embedded chips with the owner’s identifiable information. For instance, in countries such as Brazil—with a population of over 210 million people—the National ID card does not contain an embedded chip. This work explores a two-stage method, using deep learning techniques, to determine whether an ID card image provided remotely by the user is real, or tampered in the digital (composite) or non-digital domain (high-quality printed or digitally displayed on a screen). Furthermore, RGB images, frequency domain representation, noise features, and error level analysis are tested as different inputs to the two-stage classifier. The proposed BasicNet with Discrete Fourier Transform achieves the highest classification rates of 0.975 for real ID card images, and a mean of 0.968 for fake ID card images.

Faster Region-Based Hotspot Detection

As the circuit feature size continuously shrinks down, hotspot detection has become a more challenging problem in modern design for manufacturability flows. Developed deep learning techniques have recently shown their superiorities on hotspot detection tasks. However, existing hotspot detectors can only handle defect detection from one small layout clip each time, thus, may be very time-consuming when dealing with a large full-chip layout. In this article, we develop a new end-to-end framework that can detect multiple hotspots in a large region at a time and promise a better hotspot detection performance. We design a joint auto-encoder and inception module for efficient feature extraction. A two-stage classification and regression framework is designed to detect hotspot with progressive accurate localization, which provides a promising performance improvement. Experimental results show that our framework enables a significant speed improvement over existing methods with higher accuracy and fewer false alarms.

Auxiliary criterion conversion via spatiotemporal semantic encoding and feature entropy for action recognition

Video-based action recognition in realistic scenes is a core technology for human–computer interaction and smart surveillance. Although the trajectory features with the bag of visual words have confirmed promising performance, spatiotemporal interactive information cannot be effectively encoded which is valuable for classification. To address this issue, we propose a spatiotemporal semantic feature (ST-SF) and implement the conversion of it to the auxiliary criterion based on the information entropy theory. First, we present a text-based relevance analysis method to estimate the textual labels of objects most relevant to actions, which are employed to train the more targeted detectors based on the deep network. False detections are optimized by the inter-frame cooperativity and dynamic programming to construct the valid tubes. Then, we design the ST-SF to encode the interactive information, and the concept and calculation of feature entropy are defined based on the spatial distribution of ST-SFs on the training set. Finally, we achieve a two-stage classification strategy using the resulting decision gains. Experimental results on three publicly available datasets demonstrate that our method is robust and improves upon the state-of-the-art algorithms.

Application of Two-stage Classification Method Based on K-means Function Clustering Algorithm in Financial Management

Application of Two-stage Classification Method Based on K-means Function Clustering Algorithm in Financial Management

Development of an intelligent CAD system for mass detection in mammographic images

Mammography is a very useful tool to diagnose breast cancer in early stages when it is easier to treat. There are two types of evidence that radiologists look for in a mammogram, calcifications and the existence of masses. In this study, an intelligent computer-aided diagnosis system is proposed for the detection of masses in mammographic images regardless of their nature. The proposed method uses a combination of extended maxima transformations, having different threshold values, in order to find suitable internal and external markers for a marker-based watershed segmentation. After segmentation, a two-stage classifier is used to distinguish the masses better from the healthy breast tissue. A feature vector based mainly on contrast and texture features is calculated and two alternative approaches, a Bayesian classifier and a support vector machine (SVM) with Gaussian kernel function, are implemented for further reduction of the false positive areas. The system was evaluated using the data from two online databases. Specifically, 73 mammographic images from the new curated breast imaging subset of digital database for screening mammography (CBIS-DDSM) database and all the mammographic images that contain masses from the mini-mammographic image analysis society (MIAS) database were used. The overall sensitivity, in both datasets, was near 80% when the Bayesian classifier was used and above 85% when the SVM was applied.

Evaluation of methods for detecting human reads in microbial sequencing datasets.

Sequencing data from host-associated microbes can often be contaminated by the body of the investigator or research subject. Human DNA is typically removed from microbial reads either by subtractive alignment (dropping all reads that map to the human genome) or by using a read classification tool to predict those of human origin, and then discarding them. To inform best practice guidelines, we benchmarked eight alignment-based and two classification-based methods of human read detection using simulated data from 10 clinically prevalent bacteria and three viruses, into which contaminating human reads had been added. While the majority of methods successfully detected >99 % of the human reads, they were distinguishable by variance. The most precise methods, with negligible variance, were Bowtie2 and SNAP, both of which misidentified few, if any, bacterial reads (and no viral reads) as human. While correctly detecting a similar number of human reads, methods based on taxonomic classification, such as Kraken2 and Centrifuge, could misclassify bacterial reads as human, although the extent of this was species-specific. Among the most sensitive methods of human read detection was BWA, although this also made the greatest number of false positive classifications. Across all methods, the set of human reads not identified as such, although often representing <0.1 % of the total reads, were non-randomly distributed along the human genome with many originating from the repeat-rich sex chromosomes. For viral reads and longer (>300 bp) bacterial reads, the highest performing approaches were classification-based, using Kraken2 or Centrifuge. For shorter (c. 150 bp) bacterial reads, combining multiple methods of human read detection maximized the recovery of human reads from contaminated short read datasets without being compromised by false positives. A particularly high-performance approach with shorter bacterial reads was a two-stage classification using Bowtie2 followed by SNAP. Using this approach, we re-examined 11 577 publicly archived bacterial read sets for hitherto undetected human contamination. We were able to extract a sufficient number of reads to call known human SNPs, including those with clinical significance, in 6 % of the samples. These results show that phenotypically distinct human sequence is detectable in publicly archived microbial read datasets.

Energy Disaggregation Using Two-Stage Fusion of Binary Device Detectors

A data-driven methodology to improve the energy disaggregation accuracy during Non-Intrusive Load Monitoring is proposed. In detail, the method uses a two-stage classification scheme, with the first stage consisting of classification models processing the aggregated signal in parallel and each of them producing a binary device detection score, and the second stage consisting of fusion regression models for estimating the power consumption for each of the electrical appliances. The accuracy of the proposed approach was tested on three datasets—ECO (Electricity Consumption &amp; Occupancy), REDD (Reference Energy Disaggregation Data Set), and iAWE (Indian Dataset for Ambient Water and Energy)—which are available online, using four different classifiers. The presented approach improves the estimation accuracy by up to 4.1% with respect to a basic energy disaggregation architecture, while the improvement on device level was up to 10.1%. Analysis on device level showed significant improvement of power consumption estimation accuracy especially for continuous and nonlinear appliances across all evaluated datasets.

Machine Learning Classifiers Evaluation for Automatic Karyogram Generation from G-Banded Metaphase Images

This work proposes the evaluation of a set of algorithms of machine learning and the selection of the most appropriate one for the classification of segmented chromosomes images acquired using the Giemsa staining technique (G-banding). The evaluation and selection of the best classification algorithms was carried out over a dataset of 119 Q-banding chromosomes images, and the obtained results were then applied to a dataset of 24 G-band chromosomes images, manually classified by an expert of the Laboratory of Cytogenetic of the Children’s Hospital of Tamaulipas. The results of evaluation of 51 classifiers yielded that the best classification accuracy for the selected features was obtained by a backpropagation neural network. One of the main contributions of this study is the proposal of a two-stage classification scheme based on the best classifier found by the initial evaluation. In stage 1, chromosome images are classified into three major groups. In stage 2, the output of phase 1 is used as the input of a multiclass classifier. Using this scheme, 82% of the IGB bank samples and 88% of the samples of a bank of images obtained with a Q-band available in the literature consisting of 119 chromosome studies were successfully classified. The proposed work is a part of an desktop application that allows cytogeneticist to automatically generate cytogenetic reports.