Use Case: False Positives on Noise or Artifacts on Images

Overview

There may be times when a Model has multiple False Positives on noise or artifacts. This article will help you:

• Identify the signal and error
• Diagnose the root cause
• Implement the fix

Identify the Signal and Error

Your model may have difficulty differentiating between an area with noise and a defect. In this case, you will see these error signals:

Diagnose the Root Cause

The root cause is that the noise and artifacts look similar to the actual object to detect, so the Model gets confused. Look through your Train Set and see if you have sufficient samples with noise. 

False Positives on Oil Spot Defect 

Take a look at the image below. The Model falsely predicts an oil spot defect when the area is actually noise in the image.

Noisy Artifact Region 

This error occurs in the following Prediction as well.

False Positive on Punching Hole Defect 

Fix

Try one of the fixes below if you have False Positives on noise or artifacts:

• Improve your image acquisition technique or use custom data augmentation to make the training data look more like the noisy validation data. This will make your Model more robust.
• Filter out Predictions in the area that is not the region of interest by their location in the image or the pixel values.