In forensic investigations, gender identification plays a vital role in helping to identify individuals involved in criminal activities. Accurate gender identification is hindered by problems such as incomplete or degraded biological samples and limited data. The aim is to develop an accurate deep learning model of gender classification using altered low-quality images, to investigate the impact of various finger types, and to apply fingerprint reconstruction techniques. The Sokoto Coventry Fingerprint Dataset is utilized, featuring diverse fingerprint images with obliteration artificial modifications. Differences in ridge density between male and female fingerprints, with females having a higher density, have been identified as a key finding, which helps to identify the gender accurately. In demonstrating its potential for forensic use, the gender classification model achieved an excellent accuracy score of 94.84%. The classification of the finger types also shows a high accuracy of 92.39%, indicating the reliability. As demonstrated by the low mean Squared Error score and the high Structural Similarity Index score, the reconstruction of fingerprints using autoencoder models significantly improves the image quality to address practical limitations in the acquisition of clear images. These findings contribute to the development of techniques for identifying gender in forensic science, and in biometric analysis during criminal investigations. Future directions include refining feature extraction and classification models for accurate gender classification across diverse demographics, such as individuals from various countries and regions. Additionally, advancing fingerprint reconstruction techniques aims to overcome practical limitations in forensic image acquisition, enhancing overall gender classification accuracy in forensic science and biometric analysis.
