Abstract
Soft biometrics enhance the identification systems by providing supplementary information about physical or behavioral characteristics. Unlike traditional biometrics, soft biometrics focus on non-unique features that contribute valuable insights for identification, such as facial and gait characteristics. Incorporating soft biometrics improves accuracy and robustness, particularly in challenging scenarios like low-quality images or at a distant images.
This thesis introduces Soft Biometric Recognition Model (SBRM) to explore the potential of utilizing the face and gait characteristics in human recognition systems. The proposed SBRM include several stages, starting with the collection and creation of a local dataset, feature extraction where extracted three types of features: face, gait, and the combination of face and gait, preprocessing for the extracted features and finally, classification and recognition are achieved using deep learning.
The proposed model was implemented and tested using three types of features (face only, gait only, and face and gait combined). The results clearly indicate that the training accuracy score achieves around 0.98 and the testing accuracy around 0.95. The training loss value reaches approximately 0.00071 and the testing loss, which is about 0.00181 for face only. While the training accuracy value is about 0.96 and the Testing accuracy is about 0.89 with 400 Epoch. The training loss value is around 0.00179, and the testing loss value is approximately 0.00519 for gait only. The results clearly indicate that the merged features (face and gait) outperform the individual features in terms of accuracy. The merged features achieve accuracy values of 0.99 for training and 0.95 for testing. The corresponding loss values for training and testing are reported as 0.00024 and 0.00286, respectively.
