ConFERNet: a low trainable parameters based novel light-weight convolutive feature extraction recurrent network for high accuracy suspect identification

In suspect identification systems, facial features play a crucial role in recognising individuals. However, the challenge lies in sustaining the accuracy of the system over a long period of time, ensuring that it remains consistently high, reliable, and effective. This research introduces a novel lightweight model that requires low trainable parameters, a significantly smaller number than pre-trained models, which use millions of trainable parameters. The newly proposed Convolutive Feature Extraction Recurrent Network (ConFERNet) integrates a convolutional neural network and long short-term memory into a single structure to synthesise diverse images. This approach leverages computer graphics techniques to effectively extract facial features. Computer graphics play a pivotal role at various stages of this process, employing techniques such as adaptive histogram equalisation and illumination normalisation to enhance image quality under varying lighting conditions and create diverse training datasets. The LSTM-based convolutive feature-recurrent system demonstrates a notable improvement in accuracy when tested on the Augmented Reality Database (AR-DB), Extended Yale B (E-Yale B), Enhanced Extended Yale B (EE-Yale B), and Extended Cohn-Kanade (CK+) face datasets, achieving accuracy rates of 96.20%, 98.53%, 99.59%, and 99.60%, respectively. These accuracies outperform traditional baseline accuracies of 68.65% for AR-DB, 84.21% for E-Yale B, and 88.37% for CK+, suggesting the potential of this approach in enhancing suspect identification systems. This research contributes to the field by providing an innovative solution through advanced facial image feature extraction, which leads to improved accuracy rates.