How to train a network for multiple object detection.

I'm doing a cnn project of mine that would detect multiple objects in an image and I have some questions in making a cnn network and tratining and testing it too. I've read some of the cnn sample code in the website like the vehicle detector and other codes for code detection that only detect if the image is a cat/horse etc.

1. in the imageinputlayer of the CNN how many images are needed for this layer?
2. if multiple images are needed as inputs in the imageinputlayer does the image dataset(training data) going to be used in the input layer?
3. after training, what kind of images are used in the testing dataset? are they random images with different objects in it to test the detector(ex. in making a note detector should my test images be images of music sheets to test if the detector can detect multiple objects or 1picture for notes?).
4. if there are any difference in training and testing a multple object in an image(OCR) to 1 object detector on image(example stopsign detector) what are their difference?