Feature engineering applied to a probabilistic classifier in cheiloscopy

Abstract

Objectives: To design and preliminarily evaluate a feature engineering process that integrates pattern type, spatial location, and groove count as a feature vector for a probabilistic classifier aimed at human identification.

Methodology: A descriptive, non-experimental, cross-sectional study was conducted. A simple random sample of lip images (n = 100) was determined with a 95% confidence level and a precision of 3%. The units of analysis were complete lip images. Images were captured using a DSLR camera, processed with Adobe Illustrator, and exported in .JPEG format to apply a cropping mask divided into 24 cells. A coding template was used, consisting of 12 columns for the upper lip (S1 to S12) and 12 columns for the lower lip (I1 to I12). Each column was subdivided into three sections (e.g., s1, s2, s3), allowing the transcription of three biometric patterns per mask cell.

Expected Results: The obtained descriptors will be labeled and segmented. Subsequently, through Feature Engineering techniques, they will be encoded to generate a set of quantifiable characteristics that can be processed by a probabilistic classifier capable of identifying not only which pattern is present, but also its spatial location and the number of grooves that compose it. In this way, the model will enable the exploration of the discriminative capacity of the descriptors and the estimation of classification probabilities under controlled internal validation conditions. Performance will be evaluated through cross-validation, considering accuracy, sensitivity, specificity, confusion matrix, and, in verification scenarios, ROC/AUC curves and probabilistic calibration measures.