Normally captions have a separate font-setting, which should help identifying them, especially when located near to an image.

See "Parsing PDFs by text position?" and included links for a start. HTH!