Ashish Gupta

Digital Image Processing is an emerging field finding applications in different domains of science and engineering. Image processing forms the basis for pattern recognition and machine learning technologies. Color Image Processing deals with color spaces and models for performing operations on an image of one type and transform it to another model for efficient analysis and feature manipulation. In this paper, a new image enhancement method is established for better visual perception and improving image quality. Histogram of an image is useful in determining the contents in the image and its distribution. Histogram equalization is the technique used to improve the dynamic range of an image and help in distribution of parameter wrights over the entire domain range rather than it getting concentrated over specific regions. This paper proposes an effective approach to enhance image quality by histogram equalization without compromising on mean brightness aspect of the image.

According to the embodiments provided herein, a trajectory determination device for geo-localization can include one or more relative position sensors, one or more processors, and memory. The one or more processors can execute machine readable instructions to receive the relative position signals from the one or more relative position sensors. The relative position signals can be transformed into a sequence of relative trajectories. Each of the relative trajectories can include a distance and directional information indicative of a change in orientation of the trajectory determination device. A progressive topology can be created based upon the sequence of relative trajectories; this progressive topology can be compared to map data. A geolocation of the trajectory determination device can be determined.

Nuclear power plant (NPP) outages involve maintenance and repair activities of a large number of workers in limited workspaces, while having tight schedules and zero-tolerance for accidents. During an outage, thousands of workers will be working around the NPP. Extremely high outage costs and expensive delays in maintenance projects (around $1.5 million per day) require tight outage schedules (typically 20 days). In such packed workspaces, real-time human behavior monitoring is critical for ensuring safe collaboration among workers, minimal wastes of time and resources due to the lack of situational awareness, and timely project control. Current methods for detailed human behavior monitoring on construction sites rely on manual imagery data collection and analysis, which is tedious and error-prone. This paper presents a framework of automatic imagery data analysis that enables real-time detection and diagnosis of anomalous human behaviors during outages, through the integration of 4D construction simulation and object tracking algorithms.

Images have become the most popular type of mul-timedia in the Big Data era. Widely used applications like automatic CBIR underscore the importance of image understanding, especially in terms of the semantically meaningful information. Typically, high dimensional image descriptors are embedded to a subspace using a simple linear projection. However, semantic information has a complex distribution in feature space that requires a non-linear projection. We first estimate an intrinsic dimensionality of image data. Next we build a measure of visual information in embedded subspace. We compare several linear and non-linear projection methods. We use multiple image databases towards a comprehensive evaluation. We report results in terms of information content, consequent recognition rates, and computational cost. This paper is relevant for researchers interested in dimensionality reduction for large scale image understanding that is both quick and preserves semantically relevant information.

Rapidly growing technologies like autonomous navigation require accurate geo-localization in both outdoor and indoor environments. GNSS based outdoor localization has limitation of accuracy, which deteriorates in urban canyons, forested region and is unavailable indoors. Technologies like RFID, UWB, WiFi are used for indoor localization. These suffer limitations of high infrastructure costs, and signal transmission issues like multi-path, and frequent replacement of transciever batteries. We propose an alternative to localize an individual or a vehcile that is moving inside or outside a building. Instead of mobile RF transceivers, we utilize a sensor suite that includes a video camera and an inertial measurement unit. We estimate a motion trajectory of this sensor suite using Visual Odometery. Instead of pre-installed transceivers, we use GIS map for outdoors, or a BIM model for indoors. The transport layer in GIS map or navigable paths in BIM are abstracted as a graph structure. The geo-location of the mobile platform is inferred by first localizing its trajectory. We introduce an adaptive probabilistic inference approach to search for this trajectory in the entire map with no initialization information. Using an effective graph traversal spawn-and-prune strategy, we can localize the mobile platform in real-time. In comparison to other technologies, our approach requires economical sensors and the required map data is typically available in the public domain. Additionally, unlike other technologies which function exclusively indoors or outdoors, our approach functions in both  environments. We demonstrate our approach on real world examples of both indoor and outdoor locations.

Indoor positioning system is a rapidly emerging technology. Unlike outdoor positioning, which uses triangulation from satellites in line-of-sight, current indoor positioning methods attempt triangulation using Received Signal Strength Indicator (RSSI) from indoor transmitters, like WiFi and RFID. These methods, however, are not accurate and suffer from issues like multi-path and absorption by walls and other objects. In this paper we propose an alternate and novel approach to indoor positioning, that combines signals from multiple sensors. In particular, we focus on visual and inertial sensors that are ubiquitously found in mobile devices. We utilize a Building Information Model (BIM) of the indoor environment as a guideline for navigable paths. The sensor suite signals are processed to generate a trajectory of device moving through the indoor environment. We compute features on this trajectory in real-time and data mine pre-computed features on BIM's navigable paths to determine the location of the device in real-time. We demonstrate our approach on BIM in our university campus. The key benefit of our approach is that unlike previous methods that require installation of a wireless sensor network of several transmitters spanning the indoor environment, we only require a floor-plan BIM and cheap ubiquitous sensor suite on board a mobile device for indoor positioning.

This paper presents a novel adaptation of fuzzy clustering and feature encoding for image classification. Visual word ambiguity has recently been successfully modelled by kernel codebooks to provide improvement in classification performance over the standard `Bag-of-Features'(BoF) approach, which uses hard partitioning and crisp logic for assignment of features to visual words. Motivated by this progress we utilize fuzzy logic to model the ambiguity and combine it with clustering to discover fuzzy visual words. The feature descriptors of an image are encoded using the learned fuzzy membership function associated with each word. The codebook built using this fuzzy encoding technique is demonstrated to provide superior performance over BoF. We use the Gustafson-Kessel algorithm which is an improvement over Fuzzy C-Means clustering and can adapt to local distributions. We evaluate our approach on several popular datasets and demonstrate that it consistently provides superior performance to the BoF approach.

"Visual category recognition is a difficult task of significant interest to the machine learning and vision community. One of the principal hurdles is the high dimensional feature space. This paper evaluates several linear and non-linear dimensionality reduction techniques. A novel evaluation metric, the renyi entropy of the inter-vector Euclidean distance distribution, is introduced. This information theoretic measure judges the techniques on their preservation of structure in lower-dimensional sub-space. The popular dataset, Caltech-101 is utilized in the experiments. The results indicate that the techniques which preserve local neighbourhood structure performed best amongst the techniques evaluated in this paper.
"

This paper presents a novel approach to learning a codebook for visual categorization, that resolves the key issue of intra-category appearance variation found in complex real world datasets. The codebook of visual-topics (semantically equivalent descriptors) is made by grouping visual-words (syntactically equivalent descriptors) that are scattered in feature space. We analyze the joint distribution of images and visual-words using information theoretic co-clustering to discover visual-topics. Our approach is compared with the standard `Bag-of-Words' approach. The statistically significant performance improvement in all the datasets utilized (Pascal VOC 2006; VOC 2007; VOC 2010; Scene-15) establishes the efficacy of our approach.

This paper presents a novel approach to learning a visual dictionary from sub-manifolds, using co-clustering, where each sub-manifold is associated with a semantically relevant part of a visual category. The standard dictionary learning technique, called `Bag-of-Features' is limited by problems of high-dimensionality, sparsity, and noise associated with affine invariant feature descriptors. Our approach draws inspiration from the relation between object part-based models; semantic topic models; non-negative matrix factorization of multivariate data; and sub-spaces in feature space, to resolve these issues in learning a dictionary. We use co-clustering, which performs simultaneous clustering and dimensionality reduction in an optimal way, to discover multiple semantically relevant sub-spaces. We use an information-theoretic and Euclidean divergence based co-clustering. Our approach is comprehensively evaluated on several popular datasets. This work constitutes a principled first step towards a semantically
meaningful dictionary, with regards to correspondence between object parts and multiple sub-manifolds, and is not intended to compete with state-of-the-art methods like sparse coding. It is specially pertinent for the future for learning a dictionary with increasing complexity of visual categories.

In this thesis, realistic looking isolated character images indistinguishable from a writer’s individualistic writing were pseudo randomly generated by using a statistical model which learns that writer’s characteristic handwriting style. Hitherto research focus had been on modeling the human writing process or analyzing dynamic handwriting data, neither of which are viable approaches for widespread application in the growing human computer interaction technology. A writer specific statistical model of the most influential handwriting features was trained from multiple samples of each letter written by that writer from an optimal handwriting sampling text passage. Each sample letter was analyzed and new letters synthesized as a sequence of connected sub-strokes, using control-point extraction and clustering by correspondence search of multiple samples, followed by stroke curve synthesis using spline interpolation functions.

Several significant algorithms were tested for each stage of the synthesis procedure to find the techniques optimal for static handwriting data: entropy based threshold for character image extraction; Kuwahara filter for de-noising; Zhang-Suen algorithm for skeletonization; distance transform for control-point selection; shape-context descriptor for control-point correspondence search; thin plate splines for control-point transformation; and interpolating splines for generating stroke curves. Empirical results indicate that the novel combination of handwriting specific algorithms in this thesis can generate realistic synthetic handwriting in a given writer’s unique style, of satisfactory quality.

This thesis deals with the problem of estimating structure in data due to the semantic relations between data elements and leveraging this information to learn a visual model for category recognition. A visual model consists of dictionary learning, which computes a succinct representation of training data by partitioning feature space, and feature encoding, which learns a representation of each image as a combination of dictionary elements. Besides variations in lighting and pose, a key challenge of classifying a category is intra-category appearance variation. The key idea in this thesis is that feature data describing a category has latent structure due to visual content idiomatic to a category. However, popular algorithms in literature disregard this structure when computing a visual model.

Towards incorporating this structure in the learning algorithms, this thesis analyses two facets of feature data to discover relevant structure. The first is structure amongst the sub-spaces of the feature descriptor. Several sub-space embedding techniques that use global or local information to compute a projection function are analysed. A novel entropy based measure of structure in the embedded descriptors suggests that relevant structure has local extent. The second is structure amongst the partitions of feature space. Hard partitioning of feature space leads to ambiguity in feature encoding. To address this issue, novel fuzzy logic based dictionary learning and feature encoding algorithms are employed that are able to model the local feature vectors distributions and provide performance benefits.

To estimate structure amongst sub-spaces, co-clustering is used with a training descriptor data matrix to compute groups of sub-spaces. A dictionary learnt on feature vectors embedded in these multiple sub-manifolds is demonstrated to model data better than a dictionary learnt on feature vectors embedded in a single sub-manifold computed using principal components. In a similar manner, co-clustering is used with encoded feature data matrix to compute groups of dictionary elements - referred to as `topics'. A topic dictionary is demonstrated to perform better than a regular dictionary of comparable size. Both these results suggest that the groups of sub-spaces and dictionary elements have semantic relevance.

All the methods developed here have been viewed from the unifying perspective of matrix factorization, where a data matrix is decomposed to two factor matrices which are interpreted as a dictionary matrix and a co-efficient matrix. Sparse coding methods, which are currently enjoying much success, can be viewed as matrix factorization with a regularization constraint on the vectors of the dictionary or co-efficient matrices. With regards to sub-space embedding, the sparse principal component analysis is one such method that induces sparsity amongst the sub-spaces selected to represent each descriptor. Similarly, Lasso is used to induce sparsity in feature encoding by using only a sub-set of dictionary elements to represent each image. While these methods are effective, they disregard structure in the data matrix. To improve on this, structured sparse principal component analysis is used in conjunction with co-clustered groups of sub-spaces to induce sparsity at group level. The resultant structured sparse sub-
manifold dictionary is demonstrated to provide performance benefits. In a similar manner, group Lasso is used with co-clustered groups of dictionary elements to induce sparsity in terms of topics. The structured sparse encoding is demonstrated to improve aggregate performance in comparison to a regular sparse coding.

In conclusion, this thesis estimates structure in descriptor sub-spaces and learnt dictionary, uses co-clustering to compute semantically relevant sub-manifolds and topic dictionary, and finally incorporates the estimated structure in sparse coding methods, demonstrating performance gain for visual category recognition.

Learning a Structured Model for Visual Category Recognition

Abstract:
This thesis deals with the problem of estimating structure in data due to the semantic relations between data elements and leveraging this information to learn a visual model for category recognition. A visual model consists of dictionary learning, which computes a succinct representation of training data by partitioning feature space, and feature encoding, which learns a representation of each image as a combination of dictionary elements. Besides variations in lighting and pose, a key challenge of classifying a category is intra-category appearance variation. The key idea in this thesis is that feature data describing a category has latent structure due to visual content idiomatic to a category. However, popular algorithms in literature disregard this structure when computing a visual model.

Towards incorporating this structure in the learning algorithms, this thesis analyses two facets of feature data to discover relevant structure. The first is structure amongst the sub-spaces of the feature descriptor. Several sub-space embedding techniques that use global or local information to compute a projection function are analysed. A novel entropy based measure of structure in the embedded descriptors suggests that relevant structure has local extent. The second is structure amongst the partitions of feature space. Hard partitioning of feature space leads to ambiguity in feature encoding. To address this issue, novel fuzzy logic based dictionary learning and feature encoding algorithms are employed that are able to model the local feature vectors distributions and provide performance benefits.

To estimate structure amongst sub-spaces, co-clustering is used with a training descriptor data matrix to compute groups of sub-spaces. A dictionary learnt on feature vectors embedded in these multiple sub-manifolds is demonstrated to model data better than a dictionary learnt on feature vectors embedded in a single sub-manifold computed using principal components. In a similar manner, co-clustering is used with encoded feature data matrix to compute groups of dictionary elements - referred to as `topics'. A topic dictionary is demonstrated to perform better than a regular dictionary of comparable size. Both these results suggest that the groups of sub-spaces and dictionary elements have semantic relevance.

All the methods developed here have been viewed from the unifying perspective of matrix factorization, where a data matrix is decomposed to two factor matrices which are interpreted as a dictionary matrix and a co-efficient matrix. Sparse coding methods, which are currently enjoying much success, can be viewed as matrix factorization with a regularization constraint on the vectors of the dictionary or co-efficient matrices. With regards to sub-space embedding, the sparse principal component analysis is one such method that induces sparsity amongst the sub-spaces selected to represent each descriptor. Similarly, Lasso is used to induce sparsity in feature encoding by using only a sub-set of dictionary elements to represent each image. While these methods are effective, they disregard structure in the data matrix. To improve on this, structured sparse principal component analysis is used in conjunction with co-clustered groups of sub-spaces to induce sparsity at group level. The resultant structured sparse sub-manifold dictionary is demonstrated to provide performance benefits. In a similar manner, group Lasso is used with co-clustered groups of dictionary elements to induce sparsity in terms of topics. The structured sparse encoding is demonstrated to improve aggregate performance in comparison to a regular sparse coding.

In conclusion, this thesis estimates structure in descriptor sub-spaces and learnt dictionary, uses co-clustering to compute semantically relevant sub-manifolds and topic dictionary, and finally incorporates the estimated structure in sparse coding methods, demonstrating performance gain for visual category recognition.

In this thesis, realistic looking isolated character images indistinguishable from a writer’s individualistic writing were pseudo randomly generated by using a statistical model which
learns that writer’s characteristic handwriting style. Hitherto research focus had been on modeling the human writing process or analyzing dynamic handwriting data, neither of which are viable approaches for widespread application in the growing human computer interaction technology. A writer specific statistical model of the most influential handwriting features was trained from multiple samples of each letter written by that writer from an optimal handwriting sampling text passage. Each sample letter was analyzed and new letters synthesized as a sequence of connected sub-strokes, using control-point extraction and clustering by correspondence search of multiple samples, followed by stroke curve synthesis using spline interpolation functions.
Several significant algorithms were tested for each stage of the synthesis procedure to find the techniques optimal for static handwriting data: entropy based threshold for character
image extraction; Kuwahara filter for de-noising; Zhang-Suen algorithm for skeletonization; distance transform for control-point selection; shape-context descriptor for control-point correspondence search; thin plate splines for control-point transformation; and interpolating
splines for generating stroke curves. Empirical results indicate that the novel combination of handwriting specific algorithms in this thesis can generate realistic synthetic handwriting in a given writer’s unique style, of satisfactory quality.

Objective: To describe the use and experiences with external fetal monitoring devices among obstetrical providers.
Methods: Nurse, midwife, and physician obstetrical providers at an academic center were surveyed in this cross-sectional study regarding their experiences with the external fetal monitoring device utilized by their hospital system in the outpatient, inpatient, and labor and delivery (L&D) settings. The 217 providers were invited to participate between April and July, 2017. Associations between provider characteristics, device use, perception of challenging patients, and potential usefulness of an improved system were assessed with Fisher’s exact tests.
Results: Of 137 respondents, providers reported difficulties monitoring specific patient populations including obese women (98.5%), multiple gestation pregnancies (90.5%), and early gestational ages (71.5%). Most providers (84.6%) noted that patients find current devices uncomfortable. Over half (59.5%) of L&D nurses reported spending greater than 1-hour during a typical 12-hour shift interacting with these devices. Most respondents believed that an automated system would be moderately or very useful (92.6%). We found no statistically significant associations identified between provider age, years of experience, or time spent utilizing the devices with perception of challenging patient types or reporting patient discomfort.
Conclusion: This study reveals widespread views among obstetrical providers that current external fetal monitoring devices have shortcomings and that an improved system may provide benefits. These perceptions remained regardless of provider experience or time utilizing these devices. Nurses reported significant amounts of time operating the devices, representing an opportunity to reduce time and costs with an improved device.

Video analysis with the aim of discovering social relations between the people in that video is an important and unexplored topic with significant benefit towards a higher level understanding of videos. This article focuses on the inference of two social groups in each video where members of each group share friendly relations with each other and have an adversarial relation with members of the other social group. Using low-level audiovisual features and motion trajectories we compute a measure of expression of social relation in each scene in video. The occurrence of actors in each scene is computed using face recognition with LBP descriptor. The actor-scene forms a 2-model social network, which we use to compute a 1-mode network of actors. The leaders of each group, which are the actors with greater social impact are estimated using Eigencentrality. We demonstrate our approach on several Hollywood films, which span genres of action, adventure, drama, sci-fi, thriller, historic, and fantasy. This approach is successful at using video content analysis to infer the two social groups and typically the principal protagonist and antagonist in the films as well.

The task is estimating geolocation utilizing sensors and databases that reliably fun ction in GPS degraded environments and are compatible with the computing and c ommunication resources available onboard a mobile platform. Objective: Real-time geospatial localization of a mobile platform in GPS-denied e nvironment using static maps obtained from GIS database and trajectory estimatio n of onboard video camera using visual odometry. Significance: Foundational framework for coalescing information rich domains of GIS and Computer Vision technologies towards automated and ubiquitous navigati on. Approach: 3D localization of onboard camera-by computing visual features an d camera video inter-frame view geometry-is used to generate a geospatial traj ectory. This trajectory is utilized as a shape based query in a static map.