Can I use Amazon Rekognition to find cropped versions of an image?

So you want to have Amazon Rekognition Custom Labels doing Object Detection inferencing to locate specific images in a collage, where each raw image will be assigned a different label

That can be done, as long as there is a good amount of diverse examples for each label for the model training. Data augmentation techniques will help you on that, in the sense of starting from a raw image and generate variations of that by concatenating with images from different contexts. There are different libraries that help on that, but if you want to delegate that work, Amazon SageMaker Ground Truth has a brand new feature to generate synthetic data sets - more details on https://aws.amazon.com/blogs/aws/new-amazon-sagemaker-ground-truth-now-supports-synthetic-data-generation/

With an initial version of the model, you can also attach Amazon Augmented AI (Amazon A2I) to your pipeline inference so that, in cases of predictions with low score, a human reviewer can label the collage image and then can augment your training data set and train a new version of the model. That way, your model will evolve over time. One good blog post about the topic: https://aws.amazon.com/blogs/machine-learning/using-amazon-rekognition-custom-labels-and-amazon-a2i-for-detecting-pizza-slices-and-augmenting-predictions/