Why Resizing (Downsampling) Hurts Fine Detail Detection

When Resizing Might Be Necessary (and How to Mitigate Issues)

Despite the drawbacks, resizing is often a practical necessity due to:

• Computational Constraints: High-resolution images require significantly more GPU memory and processing power for training and inference. Downsampling can make training feasible on available hardware.
• Model Input Requirements: Many pre-trained CNN architectures (e.g., VGG, ResNet) expect specific input sizes (e.g., 224x224, 299x299).

Mitigation Strategies:

As per my experience Yes. Resizing alone REDUCES the performance of many models. Because when image is resized it's quality is significantly affected. If padding is followed by resizing it would be a good idea to retain the quality.

Yes. CNN are trained for fixed dimensions and the common practice is to resample and pad the input image to match the expected dimensions. Another possibility is to split the original image in tiles with some overlap, process the image tile by tile and merge the results at the end. Aerial and satellite imagery is usually very large and there is a library to standarize the tiling and merging process. It's name is SAHI. Here's the github address:

https://github.com/obss/sahi

Try to apply this library.

It is also available from Ultralytics: https://docs.ultralytics.com/es/guides/sahi-tiled-inference/

Thank you all for your responses. I completely agree with the points raised by @Mr. Antonio Ruiz and @Mr. Adnan Majid regarding the use of tiles for handling high-resolution images. This approach will be beneficial for a CNN in detecting subtle defects and assessing their severity within the images.