Metamorphic Testing of Image Captioning Systems via Reduction-based Transformations

Recently, the Image Captioning (IC) technique has been widely applied to describe a given image in text form. However, IC systems can still produce incorrect captions and lead to misunderstandings. To tackle this problem, several methods have been proposed to test the IC systems. However, these approaches still rely on pre-annotated information and hence cannot really alleviate the oracle problem in the testing. Besides, they adopt AIGC techniques to create follow-up test images that may generate unrealistic images as test cases, which leads to meaningless testing results. Thirdly, existing methods have various restrictions on the eligibility of source test cases, and hence cannot fully utilize the given images to perform testing. To tackle these issues, we propose REIC, which conducts metamorphic testing for IC systems with reduction-based transformations. Instead of relying on the pre-annotated information, we introduce a localization method to align the described objects in the caption with the corresponding objects in the test image and check whether each object in the caption retains or disappears after transformation. REIC does not artificially manipulate any objects and hence can effectively avoid generating unreal follow-up images. Besides, it eliminates the restrictions in the metamorphic transformation process, as well as decreases the ambiguity, and boosts the diversity among the follow-up test cases, which consequently enables testing to be performed on any test image, and reveals more distinct valid violations. Experimental results demonstrate that REIC can sufficiently leverage provided test images to generate follow-up cases of good reality, and effectively detect a great number of distinct violations, without the need for any pre-annotated information.