PR Curve Research Articles

Diabetes Correlated Renal Fault Prediction through Deep Learning

INTRODUCTION: Diabetic nephropathy is one of the complications of diabetes that causes damage to kidneys. Deep learning techniques are widely used to predict different diseases.OBJECTIVES: The main aim of this work is to develop an effective prediction model using deep learning. To get an effective model, a suitable dataset is considered that comprises of features related to diabetic nephropathy.METHODS: Deep belief network (DBN) is the proposed deep learning technique which is compared with naive bayes, CART decision tree, logistic regression and support vector machine. DBN is composed of Restricted Boltzmann Machines (RBM). The algorithms are analysed based on evaluation measures like area under PR curve, area under ROC curve, gini coefficient and jaccard index.RESULTS: After comparison of all algorithms, it was observed that DBN has performed better in terms of AUROC, gini coefficient and jaccard index with values 0.8203, 0.6406 and 0.7777 respectively. But CART obtained better value of 0.9039 only for AUPR.CONCLUSION: The proposed technique has outperformed other techniques in terms of three metrics and is identified as the best performing algorithm. Hence, it is suggested to use DBN while predicting diabetic nephropathy.

Development of machine learning models to predict response after cardiac resynchronization therapy

Abstract Background There have been few practical and precise tools to predict response after cardiac resynchronisation therapy (CRT). Purpose We intend to develop predictive models using machine learning (ML) approaches and easily available features prior to implantation. Methods The baseline features of 596 patients receiving CRT were retrospectively collected. Nine predictive models were established, including logistic regression (LR), Elastic Net (EN), lasso regression, ridge regression (Ridge), neural network, support vector machine (SVM), random forest, XGBoost and k-nearest neighbor. Sensitivity, specificity, precision, accuracy, F1, area under receiver operating characteristic curve (AU-ROC) and average precision of each model were evaluated, and AU-ROC was compared between each pair of ML models and further between ML models and the latest guidelines. Results Sensitivity was highest with SVM by 0.69, and specificity was highest with LR by 0.81. The models EN and Ridge showed the highest overall predictive power with an average AU-ROC of 0.77. Specifically, the Ridge model provided significant higher AU-ROC than any other model (all P<0.05). All ML models showed significant higher AU-ROC than those derived from the latest guidelines (all P<0.05). Additionally, the effect size analysis identified LBBB, LVESD, and history of PCI as the most crucial predictive features. Conclusion ML algorithms produced efficient predictive models for evaluation of response after CRT with features prior to implantation. Tools developed accordingly might improve selection of CRT candidates and reduce rate of non-response in the future. ROC and PR curves of predictive models Funding Acknowledgement Type of funding source: None

Data-Driven and Knowledge-Based Algorithms for Gene Network Reconstruction on High-Dimensional Data.

Previous efforts in gene network reconstruction have mainly focused on data-driven modeling, with little attention paid to knowledge-based approaches. Leveraging prior knowledge, however, is a promising paradigm that has been gaining momentum in network reconstruction and computational biology research communities. This paper proposes two new algorithms for reconstructing a gene network from expression profiles with and without prior knowledge in small sample and high-dimensional settings. First, using tools from the statistical estimation theory, particularly the empirical Bayesian approach, the current research estimates a covariance matrix via the shrinkage method. Second, estimated covariance matrix is employed in the penalized normal likelihood method to select the Gaussian graphical model. This formulation allows the application of prior knowledge in the covariance estimation, as well as in the Gaussian graphical model selection. Experimental results on simulated and real datasets show that, compared to state-of-the-art methods, the proposed algorithms achieve better results in terms of both PR and ROC curves. Finally, the present work applies its method on the RNA-seq data of human gastric atrophy patients, which was obtained from the EMBL-EBI database. The source codes and relevant data can be downloaded from: https://github.com/AbbaszadehO/DKGN.

Systematic Evaluation of the Diagnostic and Prognostic Significance of Competitive Endogenous RNA Networks in Prostate Cancer.

Long non-coding RNA (lncRNA)-mediated competitive endogenous RNA (ceRNA) networks act as essential mechanisms in tumor initiation and progression, but their diagnostic and prognostic significance in prostate cancer (PCa) remains poorly understood. Presently, using the RNA expression data derived from multiple independent PCa-related studies, we constructed a high confidence and PCa-specific core ceRNA network by employing three lncRNA-gene inference approaches and key node filter strategies and then established a logistic model and risk score formula to evaluate its diagnostic and prognostic values, respectively. The core ceRNA network consists of 10 nodes, all of which are significantly associated with clinical outcomes. Combination of expression of the 10 ceRNAs with a logistic model achieved AUC of ROC and PR curve up to ∼96 and 99% in excluding normal prostate samples, respectively. Additionally, a risk score formula constructed with the ceRNAs exhibited significant association with disease-free survival. More importantly, utilizing the expression of RNAs in the core ceRNA network as a molecular signature, the TCGA-PRAD cohort was divided into four novel clinically relevant subgroups with distinct expression patterns, highlighting a feasible way for improving patient stratification in the future. Overall, we constructed a PCa-specific core ceRNA network, which provides diagnostic and prognostic value.

Scene text detection using enhanced Extremal region and convolutional neural network

Text in scene images usually contains significant information. Text detection and recognition in the scene is important for a variety of advanced machine vision applications, such as image and video retrieval, automotive assistance, and multilingual translation. In particular, most text recognition systems require texts to be localized in images beforehand and this is a significant demand. The purpose of this study is to provide a method to detect texts in natural images. The proposed approach combines advantages of extremal region, ER, methods and classification of convolutional neural network, CNN. This significantly reduces the false positives and increases the accuracy of detection. The method of sliding windows is employed with different sizes in order to determine text candidates. Extraction of enhanced ERs is performed in three consecutive stages on three distinct color channels, R, G, and B. Then, the results are combined together by an add method. After grouping, the word candidates are classified to two classes of text and non-text sections by a CNN classifier. By applying non-maximum suppression (NMS) algorithm to the same words, words with the highest probability are selected. The average values of accuracy, recall, precision and F-measure of the proposed text detection model on the ICDAR2013 database are 0.893, 0.962, 0.948, and 0.955, respectively. The optimal cut point of the proposed method is 0.648, which has the highest average accuracy, 91.93%. The AUC of ROC and PR diagrams for the proposed model are 0.851 and 0.718, respectively. These results of AUC for ROC and PR curves showed an outstanding enhancement in comparison with the best detection rate of previous methods. Experimental results on the ICDAR2011, ICDAR2013 and ICDAR2015 databases also demonstrate that our algorithm outperforms the state-of-the-art scene text detection methods.

Machine Learning Based Suicide Ideation Prediction for Military Personnel

Military personnel have greater psychological stress and are at higher suicide attempt risk compared with the general population. High mental stress may cause suicide ideations which are crucially driving suicide attempts. However, traditional statistical methods could only find a moderate degree of correlation between psychological stress and suicide ideation in non-psychiatric individuals. This article utilizes machine learning techniques including logistic regression, decision tree, random forest, gradient boosting regression tree, support vector machine and multilayer perceptron to predict the presence of suicide ideation by six important psychological stress domains of the military males and females. The accuracies of all the six machine learning methods are over 98%. Among them, the multilayer perceptron and support vector machine provide the best predictions of suicide ideation approximately to 100%. As compared with the BSRS-5 score ≥7, a conventional criterion, for the presence of suicide ideation ≥1, the proposed algorithms can improve the performances of accuracy, sensitivity, specificity, precision, the AUC of ROC curve and the AUC of PR curve up to 5.7%, 35.9%, 4.6%, 65.2%, 4.3% and 53.2%, respectively; and for the presence of more severely intense suicide ideation ≥2, the improvements are 6.1%, 26.2%, 5.8%, 83.5%, 2.8% and 64.7%, respectively.

Abstract P345: Left Ventricular Hypertrophy Screening via Machine Learning for Young Adults

Background: A use of conventional 12-lead electrocardiographic (ECG) criteria and machine learning for the ECG parameters to identify the presence of left ventricular hypertrophy (LVH) is estimated 20%-30% in the general population. We developed an ECG system with simple anthropometric data using machine learning to increase the sensitivity in screening for LVH. Methods: In a large sample of 2,196 males, aged 17-45 years, the support vector machine (SVM) classifier with linear kernel is used as the machine learning method for 31 characteristics including age, body height, and body weight in addition to 28 ECG parameters such as axes, intervals and voltages to link the output of LVH. The diagnosis of LVH is based on the echocardiographic criteria for young adult males as 116 g/m 2 (left ventricular mass (LVM)/body surface area) or 49 g/m 2.7 (LVM/body height 2.7 ). The samples are divided into a training/validation set and a test set. By applying the synthetic minority over-sampling technique (SMOTE) in pre-processing of training data, the imbalance in sample size between the Non-LVH and LVH cases can be solved. Results: For the test set with 440 cases, average accuracy, specificity, sensitivity, precision and F 1 score are 76.1%, 75.1%, 92.6%, 19.5% and 32.3%, respectively. The AUCs of ROC and PR curves are 0.871 and 0.272, respectively. To compare with the Cornell and Sokolow-Lyon voltage criteria as shown in Fig.1, all data of the 2,196 young males are also tested in the model. With the specificity of 73.3%, our SVM technique provides much better sensitivity 86.7% compared to 3.3% and 52.7% for the Cornell and Sokolow-Lyon voltage criteria, respectively. The body height and body weight are the most important factors among the 31 features in our SVM model. Conclusions: We provide a novel screening tool using simple anthropometrics and several ECG data to identify LVH in young adults and improves much in sensitivity as compared with the conventional ECG criteria or using machine learning for ECG data alone to identify LVH.

Designing a Streaming Algorithm for Outlier Detection in Data Mining-An Incrementa Approach.

To design an algorithm for detecting outliers over streaming data has become an important task in many common applications, arising in areas such as fraud detections, network analysis, environment monitoring and so forth. Due to the fact that real-time data may arrive in the form of streams rather than batches, properties such as concept drift, temporal context, transiency, and uncertainty need to be considered. In addition, data processing needs to be incremental with limited memory resource, and scalable. These facts create big challenges for existing outlier detection algorithms in terms of their accuracies when they are implemented in an incremental fashion, especially in the streaming environment. To address these problems, we first propose C_KDE_WR, which uses sliding window and kernel function to process the streaming data online, and reports its results demonstrating high throughput on handling real-time streaming data, implemented in a CUDA framework on Graphics Processing Unit (GPU). We also present another algorithm, C_LOF, based on a very popular and effective outlier detection algorithm called Local Outlier Factor (LOF) which unfortunately works only on batched data. Using a novel incremental approach that compensates the drawback of high complexity in LOF, we show how to implement it in a streaming context and to obtain results in a timely manner. Like C_KDE_WR, C_LOF also employs sliding-window and statistical-summary to help making decision based on the data in the current window. It also addresses all those challenges of streaming data as addressed in C_KDE_WR. In addition, we report the comparative evaluation on the accuracy of C_KDE_WR with the state-of-the-art SOD_GPU using Precision, Recall and F-score metrics. Furthermore, a t-test is also performed to demonstrate the significance of the improvement. We further report the testing results of C_LOF on different parameter settings and drew ROC and PR curve with their area under the curve (AUC) and Average Precision (AP) values calculated respectively. Experimental results show that C_LOF can overcome the masquerading problem, which often exists in outlier detection on streaming data. We provide complexity analysis and report experiment results on the accuracy of both C_KDE_WR and C_LOF algorithms in order to evaluate their effectiveness as well as their efficiencies.

Mitral Valve Leaflet Tracking in Echocardiography using Custom Yolo3

Abstract Mitral Valve diseases still are killers in developing countries, around 80% of Cardiovascular disease patients are alone in countries who are developing. Cause of CVD is the presence of abnormalities in mitral leaflets, leaflets are key in mitral valve diseases. World Heart Federation guidelines for sonographers for disease detection of the mitral valve, the mitral leaflets’ morphology has a major role. The major drawback of these guidelines is, highly dependent on the experience of a sonographer. The main challenge is to use minimum parameters for the tracking of mitral leaflets. In this research proposed model tracked mitral leaflets to visualize more accurately, using the Yolo Mechanism with MobileNet backend of deep learning technique. It helps sonographers to identify mitral leaflets in the apical four-chamber view. We used a dataset of 40 echocardiography videos of different patients of the Apical four-chamber view(A4C). We achieved mAP of 84.4 with mitral valve leaflet detection accuracy 98% and tricuspid valve leaflet accuracy of 90% with 30 patients of 1800 image for training and 30 patients of 1800 for testing. Testing IoU 92.3%. The performance of the proposed model can be visualized by the PR curve drowned. This research will help to identify abnormalities in a mitral leaflet by visualizing them inside the boundary box.

Hierarchical infinite factor models for improving the prediction of surgical complications for geriatric patients

Nearly a third of all surgeries performed in the United States occur for patients over the age of 65; these older adults experience a higher rate of postoperative morbidity and mortality. To improve the care for these patients, we aim to identify and characterize high risk geriatric patients to send to a specialized perioperative clinic while leveraging the overall surgical population to improve learning. To this end, we develop a hierarchical infinite latent factor model (HIFM) to appropriately account for the covariance structure across subpopulations in data. We propose a novel Hierarchical Dirichlet Process shrinkage prior on the loadings matrix that flexibly captures the underlying structure of our data while sharing information across subpopulations to improve inference and prediction. The stick-breaking construction of the prior assumes an infinite number of factors and allows for each subpopulation to utilize different subsets of the factor space and select the number of factors needed to best explain the variation. We develop the model into a latent factor regression method that excels at prediction and inference of regression coefficients. Simulations validate this strong performance compared to baseline methods. We apply this work to the problem of predicting surgical complications using electronic health record data for geriatric patients and all surgical patients at Duke University Health System (DUHS). The motivating application demonstrates the improved predictive performance when using HIFM in both area under the ROC curve and area under the PR Curve while providing interpretable coefficients that may lead to actionable interventions.

CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison

Large, labeled datasets have driven deep learning methods to achieve expert-level performance on a variety of medical imaging tasks. We present CheXpert, a large dataset that contains 224,316 chest radiographs of 65,240 patients. We design a labeler to automatically detect the presence of 14 observations in radiology reports, capturing uncertainties inherent in radiograph interpretation. We investigate different approaches to using the uncertainty labels for training convolutional neural networks that output the probability of these observations given the available frontal and lateral radiographs. On a validation set of 200 chest radiographic studies which were manually annotated by 3 board-certified radiologists, we find that different uncertainty approaches are useful for different pathologies. We then evaluate our best model on a test set composed of 500 chest radiographic studies annotated by a consensus of 5 board-certified radiologists, and compare the performance of our model to that of 3 additional radiologists in the detection of 5 selected pathologies. On Cardiomegaly, Edema, and Pleural Effusion, the model ROC and PR curves lie above all 3 radiologist operating points. We release the dataset to the public as a standard benchmark to evaluate performance of chest radiograph interpretation models.

Intrusion detection using deep sparse auto-encoder and self-taught learning

With the enormous increase in the use of the Internet, secure transfer of data across networks has become a challenging task. Attackers are in continuous search of getting information from network traffic, and this is the main reason that efficient intrusion detection techniques are required to identify different kinds of network attacks. In past, various supervised and semi-supervised methods have been developed for intrusion detection. Most of these methods require a large amount of data to develop an efficient intrusion detection system. In the proposed deep neural network and adaptive self-taught-based transfer learning technique, we have exploited the concept of self-taught learning to train deep neural networks for reliable network intrusion detection. In the proposed method, a pre-trained network on regression-related task is used to extract features from NSL-KDD dataset. Original features along with extracted features from the pre-trained network are then provided as an input to the sparse auto-encoder. Self-taught learning-based extracted features, when concatenated with the original features of NSL-KDD dataset, enhance the performance of the sparse auto-encoder. Performance of self-taught learning-based approach is compared against several techniques using ten independent runs in terms of accuracy, false alarm and detection rate, area under the ROC, and PR curve. It is experimentally observed that the auto-encoder trained on the combined original and extracted features is stable and offers good generalization in comparison with the sparse auto-encoder trained on original features alone.

An efficient character recognition method using enhanced HOG for spam image detection

Generally, a spam image is an unsolicited message electronically sent to a wide group of arbitrary addresses. Due to attractiveness and more difficult detection, spam images are the most complicated type of spam. One of the ways to encounter the spam images is an optical character recognition, OCR, method. In this paper, the proposed enhanced HOG feature extraction method has been used so that the optical character recognition system of spam has been enhanced by using the HOG feature extraction method in such a way to be both resistant against the character variations on scale and translation and to be computationally cost-effective. For these purposes, two steps of the cropped image and input image size normalization have been added to pre-processing stages. Support vector machine, SVM, was employed for classification. Two heuristic modifications including thickening of the thin characters in the pre-processing stage and non-discrimination in detecting the uppercase and lowercase letters with the same shapes in the classification stage have been also proposed to increase the system recognition accuracy. In the first heuristic modification, when all pixels of the output image are empty (the character is eliminated), the original image was made thicker by one layer. In the second modification, when recognizing the letters, no differentiation was considered between the uppercase and lowercase letters with the same shapes. An average recognition accuracy of the modified HOG method with two heuristic modifications equals 91.61% on Char74K database. Then, an optimum threshold for classification was investigated by ROC curve. The optimal cutoff point was 0.736 with the highest average accuracy, 94.20%, and AUC, area under curve, for ROC and precision–recall, PR, curves were 0.96 and 0.73, respectively. The proposed method was also examined on ICDAR2003 database, and the average accuracy and its optimum using ROC curve were 82.73% and 86.01%, respectively. These results of recognition accuracy and AUC for ROC and PR curve showed an outstanding enhancement in comparison with the best recognition rate of the previous methods.

Prediction of Epitope-Associated TCR by Using Network Topological Similarity Based on Deepwalk

Currently, there are many tools available online for T-cell epitope prediction. They usually focus on the binding of peptides to major histocompatibility complex (MHC) molecules on the surface of antigen-presenting cells (APCs). However, the binding of peptides and MHC complexes to the T-cell receptor (TCR) is also critical for the immune process. Identifying the binding of human epitopes to TCRs will be useful for developing vaccines. It also has great prospects in medical issues such as cancer and autoimmune diseases. We propose a similarity-based TCR-epitope prediction method using a similarity measure. This paper introduces the Deepwalk method to calculate the topological similarity between TCR-TCRs, constructs a TCR similarity network topology, and predicts the correlation between TCRs and epitopes based on known TCR-epitope associations. We selected data from 22 types of epitopes from the VDJDB database and trained models to implement TCR-epitope prediction. We trained a model on the data from the 22 types of epitopes, predicting which epitope each TCR belongs to. To compare with other methods, we also generated a second method involving training a model for each type of epitope so that we can predict which TCR is bound to the epitope from a large pool of TCRs. We used the ROC curve, PR curve and other evaluation indicators to evaluate our model in 10-fold cross-validation. In the first model, the AUC value of our method is 0.926, and that of the support vector machine (SVM) method is 0.924. Considering that no one has ever used the first prediction model, we used the second method for the predictions. The results show better predictive performance compared to the SVM method, TCRGP method and random forest method. Our AUC values range from 0.660 to 0.950. The experimental results show that our method outperforms other methods in TCR-epitope prediction, which can help predict the TCR-epitope.

A Comparative Study of State-of-the-Art Deep Learning Algorithms for Vehicle Detection

In recent years, the deep learning object detection algorithms using 2D images have become the powerful tools for road object detection in autonomous driving. In fact, the deep learning methods for road vehicle detection have achieved the remarkable results. Although there have been a large number of studies that thoroughly explored various types of deep learning methods for vehicle detection, there are a few studies that compare and evaluate the detection time and detection accuracy of the mainstream deep learning object detection algorithms for vehicle detection. Here, this article compares five mainstream deep learning object detection algorithms in vehicle detection, namely the faster RCNN, R-FCN, SSD, RetinaNet, and YOLOv3 on the KITTI data and analyze the obtained results. The detection time and detection accuracy of the five object-detection algorithms on the KITTI test set are compared and analyzed; the PR curve and AP value are used to evaluate the detection accuracy.

Drug-Target Interaction Prediction via Dual Laplacian Graph Regularized Matrix Completion.

Drug-target interactions play an important role for biomedical drug discovery and development. However, it is expensive and time-consuming to accomplish this task by experimental determination. Therefore, developing computational techniques for drug-target interaction prediction is urgent and has practical significance. In this work, we propose an effective computational model of dual Laplacian graph regularized matrix completion, referred to as DLGRMC briefly, to infer the unknown drug-target interactions. Specifically, DLGRMC transforms the task of drug-target interaction prediction into a matrix completion problem, in which the potential interactions between drugs and targets can be obtained based on the prediction scores after the matrix completion procedure. In DLGRMC, the drug pairwise chemical structure similarities and the target pairwise genomic sequence similarities are fully exploited to serve the matrix completion by using a dual Laplacian graph regularization term; i.e., drugs with similar chemical structure are more likely to have interactions with similar targets and targets with similar genomic sequence similarity are more likely to have interactions with similar drugs. In addition, during the matrix completion process, an indicator matrix with binary values which indicates the indices of the observed drug-target interactions is deployed to preserve the experimental confirmed interactions. Furthermore, we develop an alternative iterative strategy to solve the constrained matrix completion problem based on Augmented Lagrange Multiplier algorithm. We evaluate DLGRMC on five benchmark datasets and the results show that DLGRMC outperforms several state-of-the-art approaches in terms of 10-fold cross validation based AUPR values and PR curves. In addition, case studies also demonstrate that DLGRMC can successfully predict most of the experimental validated drug-target interactions.

Improving sleep/wake classification with recurrence quantification analysis features

Abstract In this work the method of Recurrence Quantification Analysis (RQA), often used for the analysis of complex dynamic systems, is employed to extract novel features for sleep/wake classification only using cardio-respiratory signals like electrocardiogram (ECG), heart rate (HR) and respiratory effort (RE). A polysomnography data set consisting of 313 full-night recordings is used to evaluate the features. The sleep/wake classification is performed with a classifier based on Linear Discriminant Analysis (LDA) and it is validated in a Leave One Subject Out cross-validation (LOSOCV) scheme. More than 1300 features are extracted from five different cardio-respiratory modalities (ECG, RE, HR, and combinations of ECG + RE and RE + HR). Each modality is processed to obtain 13 basic RQA features, five post-processed versions and, furthermore, three normalizations are applied, leading to a total count of 975 RQA features. From literature, 126 known cardio-respiratory and actigraphy features, normalized with the same procedures thus resulting in 378 distinct features, are used for performance comparison. A feature selection method based on the Mahalanobis distance and the inter-feature correlation is used to determine the most relevant features of both sets, resulting in a set of 158 from the existing features (set A) and 232 from the RQA based features (set B). The pooled Cohen’s kappa coefficient for set A and set B is 0.586 and 0.522, respectively. The combination of both feature sets (set C) improves the kappa value to 0.625. In addition, ROC and PR curves with their corresponding Area Under Curve (AUC) values are computed. It is also shown that derived sleep statistics (sleep efficiency, sleep onset, total sleep time, etc.) deviate less with respect to the ground truth annotations using the additional RQA features compared to the use of literature-based features only. Besides the overall performance, the data set is split into six different but almost equally sized age and sex groups to allow unbiased comparisons among each other. It is shown, that the classification performance decreases with increasing age but is almost independent from sex. Hence, for future work it is suggested to implement the RQA features and also to focus on the improvement of features especially for elderly persons.

Classification assessment methods

Classification techniques have been applied to many applications in various fields of sciences. There are several ways of evaluating classification algorithms. The analysis of such metrics and its significance must be interpreted correctly for evaluating different learning algorithms. Most of these measures are scalar metrics and some of them are graphical methods. This paper introduces a detailed overview of the classification assessment measures with the aim of providing the basics of these measures and to show how it works to serve as a comprehensive source for researchers who are interested in this field. This overview starts by highlighting the definition of the confusion matrix in binary and multi-class classification problems. Many classification measures are also explained in details, and the influence of balanced and imbalanced data on each metric is presented. An illustrative example is introduced to show (1) how to calculate these measures in binary and multi-class classification problems, and (2) the robustness of some measures against balanced and imbalanced data. Moreover, some graphical measures such as Receiver operating characteristics (ROC), Precision-Recall, and Detection error trade-off (DET) curves are presented with details. Additionally, in a step-by-step approach, different numerical examples are demonstrated to explain the preprocessing steps of plotting ROC, PR, and DET curves.

Performance Evaluation of Crowd Analysis Algorithm using Modified GMM and Adaptive Thresholding

Objective: To evaluate the crowd densities in video scene under different constraints. For the crowded video analysis robust foreground detection methods are required to differentiate between moving or static foreground objects and static or dynamic background. Large number of foreground segmentation or motion segmentation approaches are available but only few can handle the various constraints like illumination variations, dynamic background partial or high level of occlusions. Method: We have proposed a modified Gaussian mixture model using adaptive thresholding. The proposed approach is implemented in MATLAB. Findings: Our proposed approach analyze all the aspects of the various backgrounds and foregrounds modelling and then compared their critical performance in terms of the PR curves and miss rate. The performance evaluation demonstrates considerable improvements in miss rate compared to traditional approaches. Our proposed method also shows significant improvements in Multi Object Detection and in Tracking Accuracy. Application: Our proposed approach analyzes the crowded scenes, especially handles outdoor environment. Optimized model parameters and adaptive thresholding makes it more robust to handle varying light conditions and partial occlusions.

Merging fixation for saliency detection in a multilayer graph

In this paper, a multilayer graph-based saliency detection algorithm by merging fixation is proposed to effectively detect salient objects in complex scenes. First, the fixation location of an image is acquired by using fixation prediction models. This is based on a motivation that human visual attention system would quickly focus on salient regions before further processing. Then, by merging the fixation saliency map into an over-segmented region map, we can obtain a coarse detection result which most likely contains salient objects. To further improve the performance of our saliency detection, the next key idea is to leverage color contrast between superpixels as features in CIE-Lab space and resolve saliency estimation of coarse regions via a multilayer graph-based framework. The final saliency detection is achieved by combining the coarse detection result with multilayer saliency maps. Extensive experiments are conducted on five benchmark datasets. Experimental results show that the proposed method yields comparable or better results in terms of PR curve, ROC curve, and F-measure, and is robust to deal with both cluttered and clean scenes.