基于Neural Networks的人脸识别：Unknowns分类与无脸判定机制问询

Great question! You’re spot-on that neural networks struggle with out-of-distribution (OOD) "Unknowns"—and face detection/recognition pipelines have tailored strategies to handle cases where no face is present, or a face isn’t recognized as part of the known set. Let’s break this down:

Face Detection: Determining if a Region Contains a Face #

• Thresholding Confidence Probabilities: This is the most common core method. Detectors like MTCNN, YOLO-Face, or RetinaFace output a confidence score (a probability between 0 and 1) for each candidate bounding box, indicating how likely the region is to contain a face. Only boxes where this score exceeds a predefined threshold (often 0.7–0.9, depending on desired precision/recall balance) are kept as valid face detections. Anything below the threshold is discarded as non-face.
• Structural Validation: Many detectors also verify the presence of key facial landmarks (eyes, nose, mouth, jawline) within the candidate box. If the landmarks don’t follow a typical human face layout (e.g., eyes are too far apart, or no nose is detected), the box is rejected—even if the confidence score is marginally above threshold. This helps filter out false positives like cartoon faces, mannequins, or random objects that resemble faces.

Face Recognition: Identifying Unknown Faces #

• Thresholding Similarity Scores: For recognition, after extracting a facial feature vector from a detected face, the system compares it to vectors of known individuals in its database (usually via cosine similarity or Euclidean distance). If the highest similarity score between the input vector and any database entry falls below a threshold, the face is classified as "Unknown"—meaning it doesn’t match any learned identity.
• Out-of-Distribution (OOD) Detection: Modern systems often add specialized OOD detection modules to complement thresholding. These might calculate how far the input face’s feature vector is from the distribution of known faces in the training set, or use an auxiliary neural network branch to explicitly predict whether the face belongs to the known category set. This reduces false acceptances of OOD faces that might accidentally have high similarity scores.

To directly address your question: Yes, thresholding predicted probabilities (or similarity scores, for recognition) is a foundational technique used in these systems. But real-world deployments almost always combine this with additional checks to improve accuracy and reduce errors.

内容的提问来源于stack exchange，提问作者John M.