Semester
Fall
Date of Graduation
2023
Document Type
Dissertation
Degree Type
PhD
College
Eberly College of Arts and Sciences
Department
Forensic and Investigative Science
Committee Chair
Jacqueline Speir
Committee Member
Casper Venter
Committee Member
Keith Morris
Committee Member
Nicholas Petraco
Committee Member
Tina Moroose
Abstract
Shoeprints deposited during the commission of a crime vary in quality as a function of numerous factors, such as substrates, media, and the specific physical activities carried out by perpetrators. This variability in impression quality impacts the amount of information available from questioned crime scene impressions, and therefore an examiner's confidence in providing opinions of source attribution. In response, this research attempted to characterize footwear impression quality based on image sharpness, contrast, entropy, homogeneity, noise, quantity of information, and complexity, and to integrate these hand-engineered feature descriptions with subjective opinions of impression quality.
The methodology followed a five-pronged approach. First, a footwear database of 597 impressions of varying quality was created. Second, automated image processing tools were employed to extract features from impressions, such as frequency and wavelet coefficients, gray-level co-occurrence matrices (GLCM), image gradients, measures of impression ratio of foreground and background, and tread contour perimeters. Third, online surveys were designed via an R-Shiny application to present image pairs and elicit quality ratings from laypersons and footwear examiners. This was followed by matrix completion for data imputation of any missing ratings. Fourth, the survey ratings and image features were examined for their suitability in ordinal (and multinomial) logistic modeling, including attributes of multicollinearity, ordinality, parallelism, and linearity. Finally, logistic regression models were created via various configurations, and the optimal model was determined via the likelihood ratio test, Akaike information criterion (AIC), and hold-one-out cross-validation (HOOCV).
A total of 49 laypersons and 31 footwear examiners responded to the R-Shiny surveys. Each was assessed for intra- and inter-rater reliability or (in)consistency. Intra-rater consistency was demonstrated by 44 out of 49 untrained (layperson) respondents, and 25 out of 31 forensic examiners. This consistency was evaluated using a cost function that described the penalty associated with changing an opinion when presented with repeated images. Next, inter-rater consistency was evaluated to determine if groups were internally calibrated. Results indicated that laypersons shared "good" agreement in quality ratings, whereas examiners shared "slight/weak" agreement in opinions. As a result, the forensic examiner dataset was subset to 123 images in an attempt to increase inter-rater reliability. Next, matrix completion was employed to infer missing ratings. The overall accuracy of matrix completion was assessed by partitioning the data into an approximate 90:10 split for training and testing. Data imputation accuracies of 96.2% and 98.6% were obtained for the layperson and examiner datasets, respectively, when allowing for no more than one rating difference between predictions and ground truth.
At this point, the examiner survey data was sequestered for future evaluation, and modeling proceeded using only the laypersons responses. Originally, an ordinal logistic regression was proposed to relate the hand-engineered feature vectors to image quality ratings. However, model assumptions of multicollinearity were severely violated, as well as modest violations of ordinality and parallelism, indicating a possible need for a multinomial logistic regression. In response, the data was transformed using principal component analysis (PCA), and both the original features as well as the PC-transformed features were regressed using both ordinal and multinomial regressions. Twenty model variations were examined, with the optimal model (based on likelihood ratio tests and AIC) resulting in 79% accuracy in predictions. This level of success is considered promising, and represents a foundation upon which continued investigation can build. The results advance the field of forensic footwear analysis by introducing an impression quality prediction model that is non-reference based, novel, and numerical. Future work will consider the relationship between layperson and examiner responses, with the goal of relating impression "quality" with "forensic information" and "weight of evidence" evaluations.
Recommended Citation
Lin, En-Tni, "Footwear Image Quality Classification: Using Subjective Assessments and Objective Image Metrics to Predict Impression Quality" (2023). Graduate Theses, Dissertations, and Problem Reports. 12254.
https://researchrepository.wvu.edu/etd/12254
Embargo Reason
Publication Pending