人脸识别模型对训练图像预测异常，请求调试排查

It’s frustrating when a model fails on an image it was trained on—let’s walk through the most likely issues and how to debug them step by step:

1. False Positives from the Face Detector #

The OpenCV DNN face detector can flag non-face regions as faces, especially with the default 0.5 confidence threshold. Here’s how to diagnose this:

• Raise the confidence threshold: Try running your script with --confidence 0.7 or 0.8. This filters out lower-confidence detections that are likely false positives.
• Log detection details: Add a print statement inside your detection loop to see the confidence and bounding box of every detected region:

  print(f"Detection {i}: Confidence = {confidence:.4f}, Bounding Box: {box.astype(int)}")
  

  This will show you if incorrect detections have low confidence values that you can filter out.
• Inspect the input image: Does your training image contain objects that look like faces (e.g., posters, reflections, toys)? The detector might be picking these up instead of (or in addition to) Aarav’s face.

2. Issues with the Recognition Pipeline #

Even if the detector finds the right face, the embedding or classification step might fail:

• Verify label encoder and recognizer match: Ensure the --recognizer and --le files were generated from the same training run. If you retrained the model but forgot to update the label encoder, you’ll get mismatched predictions.
• Check training data quality: Did you train the model with enough clear, varied images of Aarav? Were there other faces accidentally included in the training dataset? A small or noisy training set can lead to poor performance—even on training images.
• Compare embeddings: Extract the embedding of Aarav’s face from the training image during prediction, and compare it to the embeddings stored in your model. If they’re drastically different, the embedding extraction step might be the problem. Add this to your code after getting vec:

  print(f"Embedding snippet for detection {i}: {vec[0][:5]}...")  # Print first 5 values for quick comparison
  

3. Preprocessing Mismatch #

Training and prediction preprocessing must be identical—even small differences can break results:

• Check face normalization: In your code, you’re normalizing the face ROI with 1.0 / 255 when creating faceBlob. Did you use the same normalization during training? If your training code used a different scaling (e.g., no normalization), this will cause embedding mismatches.
• Detector preprocessing: The face detector uses (104.0, 177.0, 123.0) as mean values. Confirm your training pipeline used the same mean subtraction for face detection.

4. Quick Debugging Code Modifications #

To isolate the problem faster, modify your code to:

• Visualize confidence only: Temporarily change the text drawn on the image to show just confidence, not the predicted name. This lets you confirm if the detector is the issue:

  text = f"Confidence: {confidence:.2f}%"
  

• Save detected face ROIs: Write each detected face to a file so you can inspect them directly:

  cv2.imwrite(f"detected_face_{i}.png", face)
  

  If any of these saved images aren’t Aarav’s face, the detector is the culprit.
• Validate label encoder: Print out the classes in your label encoder to ensure 'Aarav' is present and correctly indexed:

  print("Label encoder classes:", le.classes_)
  

Start with the detector checks first—false positives are the most common issue here. Once you’re sure the detector is only picking up Aarav’s face, move on to verifying the embedding and recognition model.

内容的提问来源于stack exchange，提问作者Akash