How to train model on high resolution image data?

I am a beginner with convnets and am stuck up with handling of images with very high resolution of the order 3500 X 2200 and around 1 GB size. Can you please tell me how do I train my model on these images. These are basically drone imagery and also my object of interest in these images are very small (14 to 16 mm).

One efficient way is a convolutional neural network. The question is, if your node has enough ram to at least load a significant amount of data in one batch. Otherwise you have to batch sequentially.

Another, more complicated approach would be a preclustering of your image space to extract “important parts” of your image. The decision rule could depend on features like space and occupation fequency.

Another way would be to use very small windows and resample them with a network trained on the type of object (like a generative model for important object) and then use a network trained on object recognition to judge whether your window does include an important frature or not.

Important is to store the window positions as additional parameter to remap your analysis results back into the original FOV!

Good luck!!

1 Like