For contemporary law enforcement applications, especially in surveillance systems, criminal face detection and recognition are essential. Increasingly precise and effective deep learning models are required as automated systems are used increasingly frequently to identify suspects from photos or video feeds. Several cutting-edge deep learning methods for criminal face detection and recognition are examined and contrasted in this research. In particular, we examine Faster R-CNN, YOLO, SSD, and FaceNet using a proprietary criminal face database for recognition and the WIDER FACE dataset for face detection. A thorough evaluation of face detection models based on recall, accuracy, precision, and real-time inference speed is part of our research. Faster R-CNN exhibits comparatively slower processing speeds but displays better accuracy in identifying faces in difficult situations like occlusions and changing stances. On the other hand, YOLOv4 performs quite well in real-time, which makes it perfect for applications that need to quickly identify faces in live video streams. However, when occlusions are present, its performance somewhat degrades. SSD achieves a compromise between speed and accuracy, although it is not as resilient to extreme situations as Faster R-CNN. In order to match detected faces to a criminal database, we use the FaceNet model for face recognition, which produces 128-dimensional face embeddings. FaceNet's 88% recognition accuracy for criminal identification demonstrates the potential of deep learning-based recognition in practical settings. The findings show that the best detection model selection is contingent upon the particular needs of the application, such as the demand for accuracy in face detection in difficult circumstances or real-time performance. In addition to providing insights into the relative advantages of different deep learning models, this study advances the continuous development of reliable criminal face recognition systems.

Deep Learning Faster R-CNN YOLO4 SSD Face Net