Profile Face Recognition and Classification Using Multi-Task CascadedConvolutional Networks

This paper presents an in-depth analysis of Multi-task Cascaded Convolutional Networks (MTCNN) for facial recognition, focusing on both frontal and side profile face detection and classification. MTCNN, known for its ability to simultaneously perform face detection, landmark localization, and face alignment, is evaluated through a series of experiments. We assess its performance across key parameters, including cosine similarity, detection confidence, processing time, and accuracy metrics such as true positives and false positives. The results demonstrate that MTCNN excels in recognizing frontal faces with high accuracy and fast processing times, achieving excellent detection confidence and low false positive rates. While the system also performs well on side profile recognition, some challenges with false positives are observed in more difficult cases. This is estimated using a Cartesian Coordinate System along with the angles between left eye, right eye and nose in a 2D Euclidean space. It is evaluated on our own dataset as we could not find the appropriate one for this approach, with an accuracy rate of 84.7%. We also used a classifier for this approach called Random Forest Classifier. The suggested approach involves integrating our dataset into the MTCNN model, which effectively extracts key features and recognizes facial landmarks. The attributes extracted are arranged in a structured data frame with corresponding class labels, which serve as the Random Forest Classifier’s input. This classifier effectively trains to categorize and detect faces by using the extracted facial features to identify connections and patterns in the data. The application of cutting-edge deep learning techniques for feature extraction, along with the interpretability and effectiveness of the Random Forest Classifier, are important aspects of this strategy. Combining these techniques provides a dependable and expandable approach to face recognition problems, appropriate for various practical uses like biometric information identification,security, and surveillance.