A company is deploying a custom vision model using Azure Custom Vision. The training data contains images with varying resolutions. The model must achieve high accuracy. Which pre-processing step should be applied to the images before training?
Custom Vision expects consistent image sizes for optimal performance.
Why this answer
Custom Vision models use a fixed input size (e.g., 224x224 for ResNet-based architectures). Images with varying resolutions must be resized to the same dimensions before training to ensure consistent tensor shapes for the neural network. Without this step, the model cannot process the data correctly, leading to training failures or degraded accuracy.
Exam trap
Microsoft often tests the misconception that normalization or augmentation alone can compensate for varying image sizes, but the core requirement is that all images must be resized to the same dimensions to satisfy the fixed input layer of the neural network.
How to eliminate wrong answers
Option B is wrong because converting to grayscale removes color information, which is often critical for object recognition tasks (e.g., distinguishing objects by color); Custom Vision models expect 3-channel RGB input by default. Option C is wrong because pixel normalization (0-1) is typically applied internally by the Custom Vision service or as a separate step, but it does not address the fundamental requirement of uniform input dimensions. Option D is wrong because data augmentation (e.g., random cropping) is a technique to improve generalization, not a mandatory pre-processing step to handle varying resolutions; the model still requires all images to be resized to the same dimensions before augmentation.