Mean Feature Vector Research Articles

A Real-Time Memory Updating Strategy for Unsupervised Person Re-Identification

Recently, clustering-based methods have been the dominant solution for unsupervised person re-identification (ReID). Memory-based contrastive learning is widely used for its effectiveness in unsupervised representation learning. However, we find that the inaccurate cluster proxies and the momentum updating strategy do harm to the contrastive learning system. In this paper, we propose a real-time memory updating strategy (RTMem) to update the cluster centroid with a randomly sampled instance feature in the current mini-batch without momentum. Compared to the method that calculates the mean feature vectors as the cluster centroid and updating it with momentum, RTMem enables the features to be up-to-date for each cluster. Based on RTMem, we propose two contrastive losses, i.e., sample-to-instance and sample-to-cluster, to align the relationships between samples to each cluster and to all outliers not belonging to any other clusters. On the one hand, sample-to-instance loss explores the sample relationships of the whole dataset to enhance the capability of density-based clustering algorithm, which relies on similarity measurement for the instance-level images. On the other hand, with pseudo-labels generated by the density-based clustering algorithm, sample-to-cluster loss enforces the sample to be close to its cluster proxy while being far from other proxies. With the simple RTMem contrastive learning strategy, the performance of the corresponding baseline is improved by 9.3% on Market-1501 dataset. Our method consistently outperforms state-of-the-art unsupervised learning person ReID methods on three benchmark datasets. Code is made available at:https://github.com/PRIS-CV/RTMem.

Two shifts for crop mapping: Leveraging aggregate crop statistics to improve satellite-based maps in new regions

Crop type mapping at the field level is critical for a variety of applications in agricultural monitoring, and satellite imagery is becoming an increasingly abundant and useful raw input from which to create crop type maps. Still, in many regions crop type mapping with satellite data remains constrained by a scarcity of field-level crop labels for training supervised classification models. When training data is not available in one region, classifiers trained in similar regions can be transferred, but shifts in the distribution of crop types as well as transformations of the features between regions lead to reduced classification accuracy. We present a methodology that uses aggregate-level crop statistics to correct the classifier by accounting for these two types of shifts. To adjust for shifts in the crop type composition we present a scheme for properly reweighting the posterior probabilities of each class that are output by the classifier. To adjust for shifts in features we propose a method to estimate and remove linear shifts in the mean feature vector. We demonstrate that this methodology leads to substantial improvements in overall classification accuracy when using Linear Discriminant Analysis (LDA) to map crop types in Occitanie, France and in Western Province, Kenya. When using LDA as our base classifier, we found that in France our methodology led to percent reductions in misclassifications ranging from 2.8% to 42.2% (mean = 21.9%) over eleven different training departments, and in Kenya the percent reductions in misclassification were 6.6%, 28.4%, and 42.7% for three training regions. While our methodology was statistically motivated by the LDA classifier, it can be applied to any type of classifier. As an example, we demonstrate its successful application to improve a Random Forest classifier.

An improved scheme for multifeature-based foreground detection using challenging conditions

Detecting an accurate foreground from a video frame is a critical task under different complex situations such as sudden changes in illuminations, relocation of background objects, shadow, low contrast videos and dynamic backgrounds (like waving tree, rippling of water etc.). Most of the existing schemes utilize a single feature-based foreground detection approach, which in turn is hard to apply under aforementioned complex situations. In order to mitigate this issue and properly exploit the different characteristics of the pixels, the present work proposes an efficient foreground detection scheme for better segmenting the foreground. In the proposed scheme, first the texture features are first extracted utilizing cross-diagonal texture matrix (CDTM), which essentially combines the merits of both the gray level co-occurrence matrix (GLCM) and the texture spectrum (TS) to provide a complete texture information about a frame. The color and gray value features of the pixel along with texture features are utilized for the feature vector generation. Second, during background modeling phase, the similarity distance measure is computed employing the Canberra distance between the mean feature vector of the current frame and the model. Finally, a method for adaptively selecting the threshold value is proposed, instead of choosing heuristically to correctly classify the foreground and background pixels under the dynamic background condition when background pixels changing frequently.Experiments are conducted using a wide variety of indoor and outdoor video sequences under various different challenging conditions and the results compared with that of the existing state-of-the-art methods. From the experimental results, it is shown that the proposed scheme outperforms the existing schemes in terms of the quantitative as well as qualitative measures.

Autoencoder-based anomaly detection for surface defect inspection

In this paper, the unsupervised autoencoder learning for automated defect detection in manufacturing is evaluated, where only the defect-free samples are required for the model training. The loss function of a Convolutional Autoencoder (CAE) model only aims at minimizing the reconstruction errors, and makes the representative features widely spread. The proposed CAE in this study incorporates a regularization that improves the feature distribution of defect-free samples within a tight range. It makes the representative feature vectors of all training samples as close as possible to the mean feature vector so that a defect sample in the evaluation stage can generate a distinct distance from the trained center of defect-free samples. The proposed CAE model with regularizations has been tested on a variety of material surfaces, including textural and patterned surfaces in images. The experimental results reveal that the proposed CAE with regularizations significantly outperforms the conventional CAE for defect detection applications in the industry.

Developmental changes in the ability to draw distinctive features of object categories

Children draw prolifically from an early age, providing a rich source of insight into their emerging understanding of the world. Here, we investigate developmental changes in children’s ability to emphasize the relevant visual distinctions between object categories in their drawings. To do so, we developed a tablet-based platform to collect a large number of children’s drawings (5368 drawings, 1259 participants aged 2–10 years) via a free-standing drawing station in a science museum. In each session, children could produce drawings of up to 16 different object categories. On each trial, a video verbally cued a particular object category (“Can you draw a dog?”); children then had up to 30 seconds to complete their drawing. We hypothesized that drawings produced by older children would be more recognizable. To evaluate this prediction, we first applied a pretrained deep convolutional neural network model (VGG-19) to extract a high-level feature representation of all drawings. We then trained a 16-way logistic regression model under leave-one-out cross-validation to estimate the recognizability of drawings produced by children in each age group. The model’s classification accuracy increased with age, even when controlling for low-level covariates (time spent drawing, ink used, or number of strokes). This pattern replicates results from a smaller, human-rated sample of recognition judgments (N=14 raters, 286 drawings). To investigate the underlying source of these changes in recognizability, we computed the mean feature vector (center) and the mean squared distance of drawings from the category center (dispersion), for each age and category. We found both that the overall distance between category centers increased and that within-category dispersions decreased with age, suggesting that older children may have an increasing ability to include category-diagnostic information in their drawings. Future work will relate these changes in visual production ability to changes in fine motor control and object categorization.

BCIgEPRED—a Dual-Layer Approach for Predicting Linear IgE Epitopes

Allergy is a common health problem worldwide, especially food allergy. Since B cell epitopes that are recognized by the IgE antibodies act as antigenic determinants for allergy, they play a vital role in diagnostics. Hence, knowledge of an IgE binding epitope in a protein is of particular interest for identifying allergenic proteins. Though IgE epitopes may be conformational or linear, identification of the later is useful especially in food allergens that undergo processing or digestion. Very few computational tools are available for the prediction of linear IgE epitopes. Here we report a prediction system that predicts the exact linear IgE epitope. Since our earlier study on linear B-cell epitope prediction demonstrated the effectiveness of using an exact epitope dataset (in contrast to epitope containing region datasets), the dataset in this study uses only experimentally verified exact IgE, IgG, IgM and IgA epitopes. Models for Support Vector Machine (SVM) and Random Forest (RF) were constructed adopting Dipeptide Deviation from the Expected mean (DDE) feature vector. Extensive validation procedures including five-fold cross validation and two different independent dataset tests have been performed to validate the proposed method, which achieved a balanced accuracy ranging from 74 to 78% with area under receiver operator curve greater than 0.8. Performance of the proposed method was observed to be better (accuracy difference of 16–28%) in comparison to the existing available method. The proposed method is developed as a standalone tool that could be used for predicting IgE epitopes as well as to be incorporated into any allergen prediction toolhttps://github.com/brsaran/BCIgePred.

Spectral Correlative Mapping Approach for Transformation of Expressivity in Marathi Speech

The efforts of Imposing appropriate expressivity in a plain speech segment are referred to as emotion transformation. The expressivity enhances the meaning of the utterance. In speech signal, the prosody refers to emotions. This virtue is an evidence of intelligibility which humans can only add up. An appropriate prosody modification in synthetic plain Speech generated by a text to speech synthesizer plays a vital role in developing an effective speech interface platform. This research study focuses on Marathi regional language. It presents a Spectral correlative prosodic mapping approach using spectral mode decomposition (SMD) to achieve emotion transformation. The calculation of the prosodic features of the expressive speech is done by considering the segments in the signal which are high in spectral correlation. This feature vector representing the target speaking style, which is used to transplant on the mean feature vector, derived from neutral speech. Before replacing the target feature vector on the source segment its value is adaptively recomputed to reduce spectral errors between source and target emotion segments. The re-synthesized speech using prosody modified features with minimized spectral errors sounds like the target expression speech. This approach gives better quality as the computed features are not used for transformation as it is, but they are fine tuned to represent the least spectral error between source and target segments. This can also be observed by the waveform, spectrogram and objective measurements.

BCIgEPRED-a Dual-Layer Approach for Predicting Linear IgE Epitopes

Allergy is a common health problem worldwide, especially food allergy. Since B cell epitopes that are recognized by the IgE antibodies act as antigenic determinants for allergy, they play a vital role in diagnostics. Hence, knowledge of an IgE binding epitope in a protein is of particular interest for identifying aller-genic proteins. Though IgE epitopes maybe conformational or linear, identification of the later is useful especially in food allergens that undergo processing or digestion. Very few computational tools are available for the prediction of linear IgE epitopes. Here we report a prediction system that predicts the exact linear IgE epitope. Since our earlier study on linear B cell epitope prediction demonstrated the effectiveness of using an exact epitope dataset (in contrast to epitope containing region datasets), the dataset in this study uses only experimentally verified exact IgE, IgG, IgM and IgA epitopes. Models for Support Vector Machine (SVM) and Random Forest (RF) were constructed adopting Dipeptide Deviation from the Expected mean (DDE) feature vector. Extensive validation procedures including five-fold cross validation and two different independent dataset tests have been performed to validate the proposed method, which achieved a balanced accuracy ranging from 74 to 78% with area under receiver operator curve greater than 0.8. Performance of the proposed method was observed to be better (accuracy difference of 16-28%) in comparison to the existing available method. The proposed method is developed as a standalone tool that could be used for predicting IgE epitopes as well as to be incorporated into any allergen prediction toolhttps://github.com/brsaran/BCIgePred.

A Decision Support System for Classification of Normal and Medical Renal Disease Using Ultrasound Images

Early detection of medical renal disease is important as the same may lead to chronic kidney disease which is an irreversible stage. The present work proposes an efficient decision support system for detection of medical renal disease using small feature space consisting of only second order GLCM statistical features computed from raw renal ultrasound images. The GLCM mean feature vector and GLCM range feature vector are computed for inter-pixel distance d varying from 1 to 10. These texture feature vectors are combined in various ways yielding GLCM ratio feature vector, GLCM additive feature vector and GLCM concatenated feature vector. The present work explores the potential of five texture feature vectors computed using GLCM statistics exhaustively for differential diagnosis between normal and MRD images using SVM classifier. The result of the study indicates that GLCM range feature vector computed with d = 1 yields the highest overall classification accuracy of 85.7% with individual classification accuracy values of 93.3% and 77.9% for normal and MRD classes respectively.

Local learned dictionaries optimized to edge orientation for inverse halftoning.

A method is proposed for fully restoring local image structures of an unknown continuous-tone patch from an input halftoned patch with homogenously distributed dot patterns, based on a locally learned dictionary pair via feature clustering. First, many training sets consisting of paired halftone and continuous-tone patches are collected, and then histogram-of- oriented-gradient (HOG) feature vectors that describe the edge orientations are calculated from every continuous-tone patch, to group the training sets. Next, a dictionary learning algorithm is separately conducted on the categorized training sets, to obtain the halftone and continuous-tone dictionary pairs, optimized to edge-oriented patch representation. Finally, an adaptively smoothing filter is applied to the input halftone patch, to predict the HOG feature vector of an unknown continuous-tone patch, and to select one of the previously learned dictionary pairs, based on the Euclidean distance between the HOG mean feature vectors of the grouped training sets and the predicted HOG vector. In addition to using the local dictionary pairs, a patch fusion technique is used to reduce some artifacts, such as color noise and overemphasized edges on smooth regions. Experimental results show that the use of the paired dictionary selected by the local edge orientation and patch fusion technique not only reduced the artifacts in smooth regions, but also provided well expressed fine details and outlines, especially in the areas of textures, lines, and regular patterns.

Multiscale texture retrieval based on low-dimensional and rotation-invariant features of curvelet transform

Multiscale-based texture retrieval algorithms use low-dimensional feature sets in general. However, they do not have as good retrieval performances as those of the state-of-the-art techniques in the literature. The main motivation of this study is to use low-dimensional multiscale features to provide comparable retrieval performances with the state-of-the-art techniques. The proposed features of this study are low-dimensional, robust against rotation, and have better performance than the earlier multiresolution-based algorithms and the state-of-the-art techniques with low-dimensional feature sets. They are obtained through curvelet transformation and have considerably small dimensions. The rotation invariance is provided by applying a novel principal orientation alignment based on cross energies of adjacent curvelet blocks. The curvelet block pair with the highest cross energy is marked as the principle orientation, and the rest of the blocks are cycle-shifted around the principle orientation. Two separate rotation-invariant feature vectors are proposed and evaluated in this study. The first feature vector has 84 elements and contains the mean and standard deviation of curvelet blocks at each angle together with a weighting factor based on the spatial support of the curvelet coefficients. The second feature vector has 840 elements and contains the kernel density estimation (KDE) of curvelet blocks at each angle. The first and the second feature vectors are used in the classification of textures based on nearest neighbor algorithm with Euclidian and Kullback-Leibler distance measures, respectively. The proposed method is evaluated on well-known databases such as, Brodatz, TC10, TC12-t184, and TC12-horizon of Outex, UIUCTex, and KTH-TIPS. The best performance is obtained for kernel density feature vector. Mean and standard deviation feature vector also provides similar performance and has less complexity due to its smaller feature dimension. The results are reported as both precision-recall curves and classification rates and compared with the existing state-of-the-art texture retrieval techniques. It is shown through several experiments that the proposed rotation-invariant feature vectors outperform earlier multiresolution-based ones and provide comparable performances with the rest of the literature even though they have considerably small dimensions.

Comparison of organs’ shapes with geometric and Zernike 3D moments

The morphological similarity of organs is studied with feature vectors based on geometric and Zernike 3D moments. It is particularly investigated if outliers and average models can be identified. For this purpose, the relative proximity to the mean feature vector is defined, principal coordinate and clustering analyses are also performed. To study the consistency and usefulness of this approach, 17 livers and 76 hearts voxel models from several sources are considered.In the liver case, models with similar morphological feature are identified. For the limited amount of studied cases, the liver of the ICRP male voxel model is identified as a better surrogate than the female one. For hearts, the clustering analysis shows that three heart shapes represent about 80% of the morphological variations. The relative proximity and clustering analysis rather consistently identify outliers and average models.For the two cases, identification of outliers and surrogate of average models is rather robust. However, deeper classification of morphological feature is subject to caution and can only be performed after cross analysis of at least two kinds of feature vectors. Finally, the Zernike moments contain all the information needed to re-construct the studied objects and thus appear as a promising tool to derive statistical organ shapes.

Spatially correlated background subtraction, based on adaptive background maintenance

Moving object detection in dynamic backgrounds remains a challenging problem. Our earlier work established that the background subtraction using the covariance matrix descriptor is robust for dynamic backgrounds. The work proposed herein extends this approach further, using just two features-Hu moment and intensity. An improved local Hu moment is proposed, where the moment calculation of a pixel, involving neighboring pixels, are used in a weighted manner to reduce the effects of background moving pixels and the accurate shape localization of moving objects simultaneously. To further counter the erratic labeling of dynamic pixels, the fact that the neighboring pixels are spatially correlated is exploited for model construction and foreground detection. An adaptive model updating rate is calculated as a function of model distance. The proposed approach models each pixel with a covariance matrix and a mean feature vector and is dynamically updated. Extensive studies are made with the proposed technique to demonstrate its effectiveness.

Agent Searching Edge Detection Algorithm for Bone Images

Bone image segmentation is the important technology for computer aided bone diagnosis system and the foundation for three-dimensional visualization of the human skeleton. Agent searching edge detection algorithm for bone images is proposed. Based on neighbor region correlation and regional harmonic mean feature vector correlation, different species of agent accomplish searching bone edge and experimental results are satisfactory. Experimental results comparison about the proposed algorithm, Prewitt, Sobel, Log and Canny is illustrated that demonstrates the proposed algorithm has advantages in some respects.

Object Segmentation of Database Images by Dual Multiscale Morphological Reconstructions and Retrieval Applications

Processing images for specific targets on a large scale has to handle various kinds of contents with regular processing steps. To segment objects in one image, we utilized dual multiScalE Graylevel mOrphological open and close recoNstructions (SEGON) to build a background (BG) gray-level variation mesh, which can help to identify BG and object regions. It was developed from a macroscopic perspective on image BG gray levels and implemented using standard procedures, thus robustly dealing with large-scale database images. The image segmentation capability of existing methods can be exploited by the BG mesh to improve object segmentation accuracy. To evaluate the segmentation accuracy, the probability of coherent segmentation labeling, i.e., the normalized probability random index (PRI), between a computer-segmented image and the hand-labeled one is computed for comparisons. Content-based image retrieval (CBIR) was carried out to evaluate the object segmentation capability in dealing with large-scale database images. Retrieval precision-recall (PR) and rank performances, with and without SEGON, were compared. For multi-instance retrieval with shape feature, AdaBoost was used to select salient common feature elements. For color features, the histogram intersection between two scalable HSV descriptors was calculated, and the mean feature vector was used for multi-instance retrieval. The distance measure for color feature can be adapted when both positive and negative queries are provided. The normalized correlation coefficient of features among query samples was computed to integrate the similarity ranks of different features in order to perform multi-instance with multifeature query. Experiments showed that the proposed object segmentation method outperforms others by 21% in the PRI. Performing SEGON-enabled CBIR on large-scale databases also improves on the PR performance reported elsewhere by up to 42% at a recall rate of 0.5. The proposed object segmentation method can be extended to extract other image features, and new feature types can be incorporated into the algorithm to further improve the image retrieval performance.

Moving object detection in the presence of dynamic backgrounds using intensity and textural features

Moving object detection in the presence of dynamic backgrounds remains a challenging problem in video surveillance. Earlier work established that the background subtraction technique based on a covariance matrix descriptor is effective and robust for dynamic backgrounds. The work proposed herein extends this concept further, using the covariance-matrix descriptor derived from local textural properties, instead of directly computing from the local image features. The proposed approach models each pixel with a covariance matrix and a mean feature vector and the model is dynamically updated. We made extensive studies with the proposed technique to demonstrate the effectiveness of statistics on local textural properties.

Texture Based Pattern Classification

Texture can be observed in many natural and synthetic images from multispectral satellite images to the microscopic images of cell or tissue samples. Texture is an innate property of virtually all surfaces, the grain of wood, the weave of fabric, the pattern of crop in fields etc It contains important information about the structural arrangement of surfaces and their relationship to the surrounding environment. Since the textural properties of images appear to carry useful information, for discriminating purpose features have always been calculated for textures. The paper deals with the pattern classification using texture features. The features used in the project are standard deviation and entropy of all parts obtained after 3 rd level decomposition using DWT. A mean feature vector is calculated and is used for classification. For classification distance similarity is used. The different distances used are Euclidean distance, Manhattan distance, Bray Curtis distance and Canberra distance. The efficiency of classification is calculated for each distance and is compared. The time required for each retrieval, in every distance method used, is also calculated.

Unsupervised Change Detection in Satellite Images Using Principal Component Analysis and $k$-Means Clustering

In this letter, we propose a novel technique for unsupervised change detection in multitemporal satellite images using principal component analysis (PCA) and k-means clustering. The difference image is partitioned into h times h nonoverlapping blocks. S, S les h2, orthonormal eigenvectors are extracted through PCA of h times h nonoverlapping block set to create an eigenvector space. Each pixel in the difference image is represented with an S-dimensional feature vector which is the projection of h times h difference image data onto the generated eigenvector space. The change detection is achieved by partitioning the feature vector space into two clusters using k-means clustering with k = 2 and then assigning each pixel to the one of the two clusters by using the minimum Euclidean distance between the pixel's feature vector and mean feature vector of clusters. Experimental results confirm the effectiveness of the proposed approach.

On text independent speaker identification using a quadratic classifier with optimal features

In most speaker identification or verification systems it is assumed that all speakers have the same feature covariance matrices. This assumption simplifies the classification algorithm since it yields a linear classifier. Speakers, however, differ not only in their mean feature vector, but in their covariance matrix as well. The use of a speaker's individual covariance matrix results in a quadratic classifier. If a common covariance matrix is assumed, a common optimal feature space can be determined. With an individual covariance matrix, each speaker is optimally recognized in an individual feature space. The recognition scheme therefore requires the matching of an unknown speaker with the templates defined over different feature spaces. The use of the quadratic classifier together with the individual feature space is shown to drastically improve recognition accuracy while the added memory requirements are shown to be negligible. The suggested quadratic classifier, with individual optimal feature vectors, has been tested using a speaker identification system with six male speakers. In terms of a given separation measure, the quadratic classifier yields improvements of about 2 times over the conventional method.

Low-level segmentation of multispectral images via agglomerative clustering of uniform neighbourhoods

A segmentation approach based on the concept of unsupervised classification of pixels is presented. Mean feature vectors of the classes are obtained from agglomerative-type clustering of the feature values computed over uniform neighbourhoods. Two assumptions are made in this development. The first is that at least one uniform neighbourhood can be found for each of the different categories present in the image. The second is that feature vectors of neighbourhoods representative of a particular category are similar to each other, but different from those of neighbourhoods belonging to other categories. The scheme has been applied to the segmentation of a three-band multispectral image of a terrain with satisfactory results. The method is computationally efficient, and requires minimal memory; hence it can be used in real time.