In simple words: A lab result is useful only when the model works on genuinely new images.
Start with labelled images
Choose a narrow question, such as identifying damaged packages. Collect examples with reliable labels and separate training, validation, and test images. Similar photos from one object or camera session should stay in the same split to avoid an overly optimistic score.
Build and check
Resize images consistently, train a small CNN or use a suitable pre-trained model, tune with validation data, then examine test results and wrong predictions. Check whether lighting, camera angles, and rare cases change performance.
See the flow
A warehouse team labels package photos “damaged” or “not damaged”, trains a model, and checks whether it still works on photos taken on another day.
Remember: A lab result is useful only when the model works on genuinely new images.