Master of Computing and Information Systems, Youngstown State University, 2010, Department of Computer Science and Information Systems

Face sketch recognition by humans has a significant value to both criminal investigators and researchers in computer vision, face biometrics, and cognitive psychology. An important question for both law enforcement agents and scientific researchers is how accurately humans identify hand-drawn face sketches correctly. However, the experimental studies of human performance in recognizing hand-drawn face sketches are still very limited in terms of the number of artists, the number of sketches, and the number of human evaluators involved. In this study, analysis has been concluded based on psychological tests in which 406 volunteers were asked to recognize 250 sketches drawn by 5 different artists. The primary findings are: i. The sketch quality has a significant effect on human performance. Inter-artist variation as measured by the mean recognition rate can be as high as 31%. ii. Participants showed a higher tendency to match multiple sketches to one photo than to second-guess their answers. The multi-match ratio seems correlated to recognition rate, while second-guessing had no significant effect on human performance. iii. For certain highly recognized faces, their rankings were very consistent using three measuring parameters: recognition rate, multi-match ratio, and second-guess ratio, suggesting that the three parameters could provide valuable information to quantify facial distinctiveness.

Doctor of Philosophy, Case Western Reserve University, 2022, Communication Sciences

Purpose: Speech recognition in the presence of competing speech can be challenging, and individuals vary considerably in their ability to accomplish this complex auditory-cognitive task. Speech-in-speech recognition can vary due to factors that are intrinsic to the listener, such as hearing status and cognitive abilities, or due to differences in the short-term audibility of the target speech. The primary goal of the current experiments was to characterize the effects of glimpsed target audibility and intrinsic listener variables on speech-in-speech recognition. Methods: Three experiments were conducted to evaluate the effects of glimpsed target audibility, intrinsic listener variables, and acoustic-perceptual difference cues on speech-in-speech and speech-in-noise recognition. Listeners were young adults (18 to 28 years) with normal hearing. Speech recognition was measured in two stages in each experiment. In Stage 1, speech reception thresholds were measured adaptively to estimate the signal-to-noise ratio (SNR) associated with 50% correct keyword recognition for each listener in each stimulus condition. In Stage 2, keyword recognition was measured at a fixed-SNR in each stimulus condition. All participants completed a battery of cognitive measures that assessed central abilities related to masked-speech recognition. The proportion of audible target glimpses for each target+masker keyword stimulus presented in the fixed-SNR testing was measured using a computational glimpsing model of speech recognition. Results: Results demonstrated that variability in both speech-in-speech and speech-in-noise recognition depends critically on the proportion of audible target glimpses available in the target+masker mixture, even across stimuli presented at the same global SNR. Glimpsed target audibility requirements for successful speech recognition varied systematically as a function of informational masking. Young adult listeners required a greater proportion of audibl (open full item for complete abstract)

Master of Health and Human Services, Youngstown State University, 2014, Department of Health Professions

The advent of the electronic medical record (EMR) has transformed the process of clinical documentation. When combined with the speech recognition technology (SRT), EMR completeness has increased over methodologies without this technology. This research examined chart audit completion scores of physicians and scribes working within four Northeastern Ohio Emergency Services departments. SPSS® Statistics were used to perform a Repeated Measure Analysis using paired-samples t tests calculated to compare mean completion scores one month prior versus six months after SRT implementation. The mean completion score of pre-SRT implementation with and without the assistance of scribes was 5.5 ( sd = .8) and the mean completion score of post-SRT implementation without the assistance of scribes was 6.0 (sd = .9) indicating a significant increase from pre-SRT versus post-SRT implementation (t(17) = -3.9, p < 0.5). The mean completion score of pre-SRT implementation without the assistance of scribes was 5.0 (sd = 1.1) and the mean completion score of post-SRT implementation without the assistance of scribes was 6.0 (sd = .9) also indicating a significant increase from pre-SRT versus post-SRT implementation (t(17) = -4.7, p < 0.5). These analyses validated the strong statistical probability that the completeness scores of physicians utilizing SRT will exceed the total completeness scores of physicians and scribes not using this technology. Subsequently, the null hypotheses were rejected in support of the alternative hypotheses, which concluded: 1) The completeness of the EMR will at least remain the same or improve with the implementation of SRT. 2) The completeness of the EMR will at least remain the same or improve when speech recognition technology is used without scribe utilization.

Doctor of Philosophy (Ph.D.), University of Dayton, 2011, Electrical Engineering

Techniques for establishing a person7#8217;s identity are characterized by several shortcomings. Identification documents may be forged or altered, signatures are difficult to authenticate and verify, tokens or seals may be stolen or counterfeited, and physical descriptions are difficult to quantitatively assess. Establishing identity continues to be important for the same purpose it has been for centuries – for banking transactions, establishing legal presence, entering contracts, gaining entry to secured premises, identifying fugitives, etc. Recently biometrics – the science of recognizing an individual based on his physiological or behavioral traits – has gained increasing acceptance as a legitimate method for these tasks. Currently, most deployed biometric systems are unimodal – they rely on a single feature to identify a person. Although these features, such as face, iris, ear, fingerprint, signature, or voice, may be sufficiently unique, systems must still contend with a variety of problems such as noisy data, intra-class variations, restricted degrees of freedom, non-universality, spoof attacks, and unacceptable error rates. Some of these issues may be eliminated and system accuracy increased through a multimodal biometric system. In this dissertation we formulate and investigate a method for processing multimodal biometric data – collected from a single source – to extract multiple biometric features from a sample and subsequently classify the identity of the sample using multiple biometric methods in such a way that some or all of the identity features may be opportunistically selected. In this context, opportunistic selection is meant to refer to techniques used to identify sporadic but consistently unavailable features which may be used to improve the classification rate of a more reliably present feature. Specifically, we work with three biometrics extracted from a single high-resolution near infrared face image: iris, face, and skin irregularity featur (open full item for complete abstract)

Master of Science, The Ohio State University, 2005, Graduate School

Master of Science, Miami University, 2023, Computer Science and Software Engineering

Software models are used to analyze and understand the properties of the system, providing stakeholders with an overview of how the system should work before actually implementing it. Such models are usually created informally, such as drawing sketches on a whiteboard or paper, especially during the early design phase, because these methods foster communication and collaboration among stakeholders. However, these informal sketches must be formalized to be useful in later applications, such as analysis, code generation, and documentation. This formalization process is often tedious, error-prone, and time-consuming. In an effort to avoid recreating formal models from scratch, this thesis presents SkeMo, a sketch-based software modeling tool. SkeMo is built on a CNN-based image classifier using 3000 input sketches of class diagram components and integrated into the functionality of an existing web-based model editor, the Instructional Modeling Language (IML), with a newly implemented touch interface. SkeMo was evaluated using a ten-fold cross-validation to assess the image classifier and through a user study involving 20 participants to collect metrics and feedback. The results demonstrate the promising potential of sketch-based modeling as an intuitive and efficient modeling practice, allowing users to quickly and easily create models to design complex software systems.

Master of Science (M.S.), Xavier University, 2023, Psychology

Employee wellbeing is an important construct that can affect several organization-wide outcomes. The purpose of this study was to examine the effect of one potential predictor of wellbeing, namely recognition. This study also examined if employee engagement moderated the relationship between recognition and wellbeing. It was hypothesized that participants in the recognition conditions would report significantly higher levels of wellbeing than those in the control condition. It was also hypothesized that engagement would moderate the relationship between recognition and wellbeing, such that this relationship would be stronger for participants high on engagement. Participants were first given a measure of engagement and were then randomly assigned to one of four scripts depicting a conversation with a hypothetical manager in which they were received one of four conditions of recognition: two forms of recognition (acknowledgement and gratitude), only acknowledgement, only gratitude, or neither form of recognition. Then, they completed a measure of wellbeing based on the interaction they read about with their hypothetical manager, as well as a demographics form. Finally, they were debriefed. Results showed that participants who received recognition did not report significantly higher wellbeing than those who did not receive recognition, and that engagement did not significantly moderate this relationship. However, engagement was positively and significantly related to wellbeing. These findings suggest that a simple “thank you” or “good job” statement may not be enough to enhance wellbeing, and that managers should focus on increasing employee engagement instead. Nevertheless, future research should be conducted to re-examine these relationships using employees in an actual organization that has an existing recognition program.

Master of Computer Science (M.C.S.), University of Dayton, 2022, Computer Science

A picture is worth a thousand words. A single image has the power to influence individuals and change their behaviour, whereas a single word does not. Even a barely visible image, displayed on a screen for only a few milliseconds, appears to be capable of changing one's behaviour. In this thesis, we experimentally investigated the relationship between facial expressions and perceived emotions. To this end, we built two datasets, namely, the image dataset for image emotion analysis and the face dataset for expression recognition. During the annotation of the image dataset, both facial expressions and perceived emotions are recorded via a mobile application. We then use a classifier trained on the face dataset to recognize the user's expression and compare it with the perceived emotion.

Doctor of Philosophy, The Ohio State University, 2022, Speech and Hearing Science

The three manuscripts presented here examine concepts related to speech perception in noise and ways to overcome poor speech intelligibility without depriving listeners of environmental sound recognition. Because of hearing-impaired (HI) listeners' auditory deficits, there is a substantial need for speech-enhancement (noise reduction) technology. Recent advancements in deep learning have resulted in algorithms that significantly improve the intelligibility of speech in noise, but in order to be suitable for real-world applications such as hearing aids and cochlear implants, these algorithms must be causal, talker independent, corpus independent, and noise independent. Manuscript 1 involves human-subjects testing of a novel, time-domain-based algorithm that fulfills these fundamental requirements. Algorithm processing resulted in significant intelligibility improvements for both HI and normal-hearing (NH) listener groups in each signal-to-noise ratio (SNR) and noise type tested. In Manuscript 2, the range of speech-to-background ratios (SBRs) over which NH and HI listeners can accurately perform both speech and environmental recognition was determined. Separate groups of NH listeners were tested in conditions of selective and divided attention. A single group of HI listeners was tested in the divided attention experiment. Psychometric functions were generated for each listener group and task type. It was found that both NH and HI listeners are capable of high speech intelligibility and high environmental sound recognition over a range of speech-to-background ratios. The range and location of optimal speech-to-background ratios differed across NH and HI listeners. The optimal speech-to-background ratio also depended on the type of environmental sound present. Conventional deep-learning algorithms for speech enhancement target maximum intelligibly by removing as much noise as possible while maintaining the essential characteristics of the target speech signal (open full item for complete abstract)

PHD, Kent State University, 2022, College of Arts and Sciences / Department of Psychological Sciences

Research indicates that readers break down complex words into their smallest, meaning-based units (morphemes) when spelling (e.g., Senechal, 2000). However, it remains unclear how morphemes are formed and whether newly formed morphological knowledge (i.e., knowledge of morphemes) is strong enough to support word learning. The current research proposes to address this gap in the literature by investigating how adult readers use recently acquired morphological knowledge when learning compound words. The first morpheme in each of the compound words was a novel non-word. Participants learned the meanings of the novel morphemes (e.g., breese = “fish”), and after a one-day delay, participants were re-exposed to those morphemes in novel compound words that contained the novel morpheme that had been learned the previous day (e.g., breesebin). The compound words were presented in sentence context that taught participants meanings to the compound words that either (1) were consistent with the meaning of the novel morpheme contained in the compound word (e.g., breesebin = “fish storage”); (2) were inconsistent with the meaning of the novel morpheme contained in the compound word (e.g., breesebin = “alleyway”); or (3) contained no contextual cues from which the meaning of the compound word could be derived. The current research found that participants were more successful at learning the novel compound words when they were able to use morphological knowledge from the novel word learned on day one to support their learning on day two. This was true for both learning the spellings of words and learning the meanings of words, despite participants receiving no instructions to consider the words' morphological knowledge. These findings support the body of literature that stress the importance of emphasizing morphology during language instruction.

Doctor of Philosophy, The Ohio State University, 2021, Computer Science and Engineering

The last couple of decades have seen vast growth in the adoption of speech as an input modality, spread across a variety of tasks and devices. There has been corresponding rapid growth in the use of Automatic Speech Recognition (ASR) systems to transcribe the audio into text. ASR technology has advanced over time, however even today the best speech recognition systems still make errors, and even a single word error in recognizing an utterance can change the semantics or understanding of an utterance. The proliferation of automatic Natural Language Understanding (NLU) has also grown through the use of chatbots, machine translation, topic modeling, information retrieval, language modeling, and many more applications. Due to the natural style of speech, developers that incorporate NLU technology in their applications have increasingly felt a need to provide for the acceptance of speech input in replacement or in addition to typed text. A popular approach has been to utilize a cloud-based ASR service as a front-end to convert speech into text before feeding it to their text-based NLU back-end. Until developers deploy NLU systems in the spoken domain for long enough to collect adequate labeled data, they end up needing to rely on plain text data i.e., typed or otherwise devoid of ASR errors, to train their systems. When an NLU system trained on plain text is deployed under speech input, the presence of ASR errors changes the intended spoken text before it reaches the NLU system affecting downstream performance. Such a lack of ASR error examples in the plain text training data leads to a missed opportunity for NLU systems to become robust to the errorful input they may see at test-time. ASR error prediction is a technique to take plain text and predict the kinds of errors that may have occurred if that text were to have been spoken and transcribed by an ASR system. When plain text data is to be used to train systems for spoken language understanding or ASR, a proven st (open full item for complete abstract)

MS, University of Cincinnati, 2020, Engineering and Applied Science: Computer Science

Classification of specialized human activity datasets utilizing methods not requiring manual feature extraction is an underserved area of research in the field of human activity recognition (HAR). In this thesis, we present a convolutional neural network (CNN)-based method to classify a dataset consisting of subjects lifting an object from various positions relative to their bodies, labeled by the level of back pain risk attributed to the action. Specific improvements over other CNN-based models for both general and activity-based purposes include the use of average pooling and dropout layers. Methods to reshape accelerometer and gyroscope sensor data are also presented to encourage the model's use with other datasets. When developing the model, a dataset previously developed by the National Institute for Occupational Safety and Health (NIOSH) was used. It consists of 720 total trials of accelerometer and gyroscope data from subjects lifting an object at various relative distances from the body. In testing, 90.6% accuracy was achieved on the NIOSH lifting dataset, a significant improvement over other models tested. Saliency results are also presented to investigate underlying feature extraction and justify the results collected.

Bachelor of Science (BS), Ohio University, 2019, Communication Sciences and Disorders

People with high noise exposure are more at risk of developing a hearing loss. Some of these high-risk people have normal audiograms but have problems hearing in different environments, such as noisy restaurants. Cochlear synaptopathy is when there is specific damage to afferent auditory nerve synapses onto the inner hair cells. This loss is considered “hidden” because its inimical effects are often not evident on conventional hearing tests. This study aims to examine: 1) the extent that noise exposure history affects the amplitude of auditory brain response (ABR) wave I and 2) what the influence of noise exposure history has on speech in noise performance. A noise exposure questionnaire was given to people interested in participating. During session 1, otoscopy, tympanometry, a conventional hearing screening, a high-frequency hearing screening, the CNC word test, and the AzBio sentence test were done. During session 2, distortion product otoacoustic emissions (DPOAE) and ABR tests were done. The results showed that there was a significant difference in performance on the consonant-nucleus-consonant (CNC) word task and AzBio sentence task between the two noise exposure background (NEB) groups. There was not a significant correlation in the ABR results between the two NEB groups.

Master of Science in Electrical Engineering (MSEE), Wright State University, 2016, Electrical Engineering

The application of Automatic Target Recognition (ATR) on High Range Resolution (HRR) radar data in a scenario that contains unknown targets is of great interest for military and civilian applications. HRR radar data provides greater resolution of a target as well as the ability to perform ATR on a moving target, which gives it an advantage over other imaging systems. With the added resolution of HRR comes the disadvantage that a change in the aspect angle or orientation results in greater changes in the collected data, making classical ATR more difficult. Closed set ATR on HRR radar data is defined when all potential targets are assumed to be part of the training target data base. Closed set ATR has been able to achieve higher rates of correct classification by the selection of proper feature extraction algorithms, however, only a few methods for performing open set ATR have been developed. Open set ATR is the ability to identify and discard when a target is not one of the trained targets. By identifying these untrained targets, the number of misclassified targets is reduced, thereby, increasing the probability of a correct classification in a realistic setting. While the open set ATR produces a more realistic approach, the classical closed-set ATR is the standard method of ATR. One of the more popular classification algorithms currently used today is the Support Vector Machine (SVM). The SVM by nature only works on a binary closed-set problem. However, by extracting probabilities from an SVM as proposed by Platt [1], this classification algorithm can be applied to open set. In this thesis, the feature extraction methods established in closed-set ATR are modified to facilitate the application of the Probabilistic Open Set Support Vector Machine (POS-SVM). Utilizing the Eigen Template (ET) and Mean Template (MT) feature extraction methods developed for closed-set ATR, in combination with centroid alignment, an open set ATR Probability of correct classification (PCC (open full item for complete abstract)

Master of Science in Engineering (MSEgr), Wright State University, 2014, Electrical Engineering

Classification algorithms trained using finite sets of target and confuser data are limited by the training set. These algorithms are trained under closed set assumptions and do not account for the infinite universe of confusers found in practice. In contrast, classification algorithms developed under open set assumptions label inputs not present in the training data as unknown instead of assigning the most likely class. We present an approach to open set recognition, the probabilistic open set SVM, that utilizes class posterior estimates to determine probability thresholds for classification. This is accomplished by first training an SVM in a 1-vs-all configuration on a training dataset containing only target classes. A validation set containing only class data belonging to the training set is used to iteratively determine appropriate posterior probability thresholds for each target class. The testing dataset, which contains targets present in the training data as well as several confuser classes, is first classified by the 1-vs-all SVM. If the estimated posterior for an input falls below the threshold, the target is labeled as unknown. Otherwise, it is labeled with the class resulting from the SVM decision. We apply our method to classification of synthetic ladar range images of civilian vehicles and measured infrared images of military vehicles. We show that the POS-SVM offers improved performance over other open set algorithms by allowing the use of nonlinear kernels, incorporating intuitive free parameters, and empirically determining good thresholds.

MS, University of Cincinnati, 2014, Engineering and Applied Science: Electrical Engineering

Biometric technology is playing a vital role in the present day due to the rapid development of secure systems and home automation that have made our lives easier. But the question arises as to how far these systems are secure. With advances in hacking, the traditional username and password security protocol is not optimal for all security based systems. Though fingerprint identification systems provided a path-breaking solution, there are many methods to forge fingerprints. While other technologies like voice recognition, iris recognition, etc., co-exist, the security and safety of these technologies are also open to question. The major objective of this thesis is to provide enhanced security through a biometrics based embedded system using the technique of Vascular Pattern Recognition or Vein Pattern Recognition (VPR). Another objective is to enhance the vascular pattern image through various image processing techniques. Another target is to reduce the Comparison for Result (CFR) time by a significant factor. Finally, the aim is to implement this VPR based embedded system in a real time software environment. For the system we implemented, our experiments achieved a false accept rate of 0% and a false reject rate of 6.34%. Furthermore, it has been demonstrated in our research that the Speeded Up Robust Features (SURF) algorithm is faster than its predecessor algorithm Scale Invariant Feature Transform (SIFT). The principal conclusion of the thesis is that a safe and secure system can be developed on a small scale with precise results. Given the resources, this system could be extended to a larger scale and customized for a wide range of applications.

Doctor of Philosophy, The Ohio State University, 2014, Psychology

A powerful theoretical framework for exploring recognition memory is the global matching framework, in which a cue's memory strength reflects the similarity of the retrieval cues being matched against all of the contents of memory simultaneously. Contributions at retrieval can be categorized as matches and mismatches to the item and context cues, including the self match (match on item and context), item noise (match on context, mismatch on item), context noise (match on item, mismatch on context), and background noise (mismatch on item and context). I present a model that directly parameterizes the matches and mismatches to the item and context cues, which enables estimation of the magnitude of each interference contribution (item noise, context noise, and background noise). The model was fit within a hierarchical Bayesian framework to ten recognition memory datasets that employ manipulations of strength, list length, list strength, word frequency, study-test delay, and stimulus class in item and associative recognition. Estimates of the model parameters revealed at most a small contribution of item noise that varies by stimulus class, with virtually no item noise for single words and scenes. Despite the unpopularity of background noise in recognition memory models, background noise estimates dominated at retrieval across nearly all stimulus classes with the exception of high frequency words, which exhibited equivalent levels of context noise and background noise. These parameter estimates suggest that the majority of interference in recognition memory stems from experiences acquired prior to the learning episode.

Master of Computing and Information Systems, Youngstown State University, 2007, Department of Computer Science and Information Systems

Face recognition has received much attention recently in the biometrics research. Many studies have shown improvement in recognition rate when 2D and 3D faces were combined. However, the use of the 3D face has a few limitations such as that the 3D data requires much more storage space and long processing time. Therefore, there is a strong interest to explore new methods that can provide similar or better results in the face recognition. This thesis presents an experimental study by using a sequence of rotating head videos under two different lighting conditions, regular indoor lighting and strong shadow lighting. The experiment were carried out using two sets of data, the first set of over 100 subjects and the second set of 47 subjects. Very promising results have been observed in terms of the recognition performance measured by the cumulative characteristics curves.

Doctor of Philosophy (PhD), Wright State University, 2009, Engineering PhD

We have established a modular virtual framework to design accurate, robust, efficient and cost-conscious autonomous target/object detection systems. Developed primarily for image-based detection problems, such as automatic target detection or computer-aided diagnosis, our approach is equally suitable for non-image-based pattern recognition problems. The framework features six modules: 1) the detection algorithm module accepts two-dimensional, spatially-coded sensor outputs; 2) the evaluation module uses our receiver operator characteristic (ROC)-like assessment tool to evaluate and fine-tune algorithm outputs; 3) the fusion module compares outputs combined under various fusion schemes; 4) the classifier selection module exploits the double-fault diversity measure (F2 DM) to identify the best classifier; 5) the weighting module judiciously weights the algorithm outputs to fine-tune classifiers, and 6) the cost-function analysis module determines the best detection parameters based on the trade-off between the costs of missed targets and false positive detections. Our solution can be generalized to facilitate detection system design in various applications, including target detection, medical diagnosis, biometrics, surveillance, machine vision, etc. For proof-of-principle, the framework was implemented for the autonomous detection of roadside improvised explosive devices (IEDs). From our set of nine multimodal detection algorithms that yield 1,536 possible classifiers, we identified the single best classifier to accomplish the detection task under a defined cost specification. System performance was tracked through each module and compared to standard approaches for system definition. Algorithm parameter optimization improved performance by an average of 18% (range of 3-32%). Our F2 DM-based classifier selection module predicted classifier performance with an average difference of 3% (standard deviation = ± 2%) from ROC area under the curve (AUC) predictions and an as (open full item for complete abstract)

Doctor of Philosophy, The Ohio State University, 2009, Electrical and Computer Engineering

In American Sign Language (ASL), the manual and the non-manual components play crucial semantical and grammatical roles. The design of systems that can analyze and recognize ASL sentences requires the recovery of both these manual and non-manual components. Manual signs in ASL are constructed using three building blocks – handshape, motion, and place of articulation. Only when these three are successfully estimated, can a sign be uniquely identified. The first part of my research is to define algorithms to recognize manual signs based on the recovery of these three components from a single video sequence of two-dimensional images of a sign. The 3D handshape is obtained with a structure-from-motion algorithm based on the linear fitting of matrices with missing data. To recover the 3D motion of the hand, a robust algorithm is defined which selects the most stable solution from the pool of all the solutions given by the three point resection problem. Faces of the signers in the video sequence are detected, with which the coordinate system with respect to the signer is defined and hence we recover the place of articulation of the sign. Based on the recognition results of the three recovered components, the manual signs are recognized using a tree-like structure. For the non-manual component of ASL, we need to provide an accurate and detailed description of external and internal facial features. The second part of this research focuses on the precise detailed detection of faces and facial features. Learning to discriminate the features from their context permits a precise detection of facial components, which is the key point of the feature detection algorithm. And because the shape and texture of facial features vary widely under changing expression, pose and illumination, the detection of a feature versus the context is challenging. This problem is addressed with the use of subclass division, which is employed to automatically divide the training samples of each facial (open full item for complete abstract)