How to Pick the Optimal Image Size for Training Convolution Neural Network?
When training a Convolution Neural Network on a custom dataset, picking the right image is crucial. This will impact the training time & performance of the model. Also, we will learn how to identify if there are any issues with the dataset.
Why?
Before we jump to how part, Let us discuss the negative consequence of choosing the wrong size. After we pick a fixed width and height, the standard procedure is to resize all the images to this fixed size. So, now every image falls into one of the two buckets.
- Downscaling: Bigger images will be down scaled, this makes it harder for CNN to learn the features required for classification or detection as the number of pixels where the vital feature will be present is significantly reduced.
- Upscaling: When small images are upscaled and padded with zero, then NN has to learn that the padded portion has no impact on classification. Larger images are also slower to train and might require more VRAM.
So we have to pick our poison, the closer to optimal image size we are, the better it is.
Optimal Image Size
We all know that choosing the right size depends on the dataset, but the question is how to do it? Visualize the image size.
Dataset: The Oxford-IIIT Pet Dataset
This dataset has more than 7000 images with varying size and resolution.
From the first plot, it looks like most images are of resolution less than 500 by 500. After zooming in, we can clearly see that images are clustered around either size 300 or 500. My recommendation for this dataset is to start training the neural network with image size 300 and progressively increase it to 400 and finish it with size 500. By this way, the model should be able to generalize well for different image resolutions.
Bonus
- Wouldn’t it be great if we can see the underlying images while inspecting the points in the plot? This can help us to identify the following potential issues.
- Mislabeled data: Can confuse the model.
- Certain class images are of very high or low resolution: Could make the model biased.
2. After inspection, if we conclude that some data points should be removed, then use Lasso Selector to achieve that.
