Use Case: Loss of Details from Low-Quality Images

Overview

There may be times when a Model has low performance due to low-quality images. This article will help you:

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

Identify the Signal and Error

If your Model has several False Negatives due to low-quality images, this is an indicator that the Model is not performing well on images with low detail. These images could be overexposed, underexposed, blurry, or could have lost detail due to the imaging solutions.

Diagnose the Root Cause

You can diagnose the root cause by parsing through your training data. For this use case of losing image quality, there is generally only one root cause: the image quality in the Train Set is too low for the Model to detect the objects.

This use case will likely cause several False Negatives, like in the image below.

The Model Failed to Detect an Object Because the Image Resolution Is Too Low 

Fix

Now that you've learned how to identify the signal and diagnose the cause, you can iterate your data to achieve higher accuracy by removing low-quality images. If this is not possible, try adding random blur augmentations to make the Model more robust to blurry images. 

Apply These Tips to This Stage of the Workflow