<p>Edge detection in inkjet-printed images is fundamental to achieving high-precision quality control; however, it is confronted by three major physical challenges: edge blurring and ink bleed, substrate texture noise, and multi-colour registration errors. These challenges not only test the limits of both conventional and deep learning algorithms but also highlight a key inadequacy in current evaluation frameworks. General academic metrics, such as the F1-score, focus on visual fidelity but often do not sufficiently quantify an algorithm’s efficacy in addressing real-world industrial problems, such as measuring edge roughness or registration error. This points to a notable gap between research and practical application. The core contribution of this paper is, therefore, to systematically review state-of-the-art techniques while concurrently proposing and substantiating the necessity of establishing a framework for ’industrial quantitative evaluation’. This review advocates for a paradigm shift from ’visual correspondence’ to ’physical quantification’. This paper first analyses the limitations of conventional operators and subsequently provides a focused discussion on how deep learning methods, centred on Convolutional Neural Networks (CNNs), address the aforementioned challenges through multi-scale feature learning and contextual modelling. Finally, the paper offers an outlook on frontier directions, including physics-informed models and self-supervised learning. Through a dual-path approach of ’technological scrutiny’ and ’paradigm reshaping’, it aims to provide a theoretical reference and a technical roadmap both in-depth and forward-looking–for quality control and process optimisation within the field of inkjet printing.</p>