The study, published as an open-access article in the journal Multimedia Tools and Applications, was conducted according to the PRISMA guidelines for systematic reviews, the same rigorous protocol used in medical meta-analyses. The researchers searched the Scopus database along with the proceedings of the field’s flagship venues, including CVPR, NeurIPS, ICLR, KDD, ECCV and ICCV, covering publications from January 2014 to July 2025. The initial sweep returned 2,788 candidate publications. After a three-stage screening process that removed duplicates, irrelevant records and studies of insufficient quality, 286 high-quality studies remained, forming the largest and most transparent evidence base yet assembled on knowledge-guided approaches to computer vision.

The motivation behind the survey stems from a growing unease in the AI community about what happens when models rely on data alone. The authors point to well-documented failures of this philosophy, including training on the ImageNet dataset producing models that learned a distorted picture of the world, for instance associating chief executives with white men and flight attendants with Asian women. Beyond bias, purely data-driven models are vulnerable to adversarial perturbation, can leak sensitive information memorized during training, become overconfident when deployed on incomplete data, and often fail to generalize beyond their training distributions. As governments worldwide impose stricter regulations on artificial intelligence, these shortcomings have shifted from academic curiosities to practical obstacles.

KGML offers a way out by embedding structured knowledge directly into the learning process, and the survey draws a sharp conceptual line between two kinds of that knowledge. Scientific knowledge is grounded in established disciplines: physical laws such as illumination and reflectivity, geometric principles governing shape and spatial configuration, semantic relationships between objects, and domain expertise from medicine, biology, environmental science and mathematics. Non-scientific knowledge, by contrast, refers to auxiliary information drawn from the machine learning ecosystem itself, including knowledge graphs, knowledge distillation, knowledge transfer and generative learning. A medical imaging model constrained by anatomical structure falls into the first category; a detector that consults a knowledge graph encoding that a cat typically sits on a table falls into the second.

The authors further organize how knowledge is actually fused into models, adapting a framework originally proposed for scientific machine learning into three integration strategies. Knowledge-guided learning embeds domain principles into loss functions during training, so that a model is penalized not only for prediction errors but for violating known physics or rules. Knowledge-guided architecture hard-codes constraints directly into network structure, as in physics-guided neural networks for lake temperature prediction, where physical restrictions are built into the topology itself. Knowledge-guided pre-training uses scientific or simulated data to initialize model parameters, reducing how much real-world labeled data is needed to fine-tune the system. The survey shows all three strategies have flourished across the field’s major tasks, including classification, object detection, segmentation and multimodal understanding.

The reviewed literature reveals strikingly diverse applications of scientific knowledge in vision. Environmental and earth science applications range from monitoring forests in the Swiss Alps with multi-temporal deep learning guided by knowledge of forest cover dynamics, to mapping land surface temperature, detecting natural gas micro-leaks from hyperspectral imagery and recognizing mineral grains. In biology, hierarchy-guided neural networks have improved fish species classification, particularly when training data is scarce, by exploiting phylogenetic relationships. In medicine, KGML underpins ultrasound elastography, choroid segmentation in retinal images, medical report generation and facial pain assessment in intensive care units. Mathematical knowledge has been harnessed too, with systems like FOMAS splitting reasoning into a knowledge system that learns formulas and a reasoning system that applies them, echoing dual-process theories of human cognition.

On the non-scientific side, the survey dissects four dominant mechanisms. Knowledge graphs provide structured semantic context that sharpens object detection and scene understanding by encoding relationships among entities. Knowledge distillation transfers what a large, accurate teacher network knows into a compact student model, often through soft probability labels that carry richer information than hard ones; relation-based variants go further, teaching students to reproduce the structural relationships among a teacher’s internal representations. Knowledge transfer and knowledge bases allow models trained on one task or domain to bootstrap performance on related problems with little data, a strategy central to few-shot and long-tail recognition. Generative learning, meanwhile, uses prior knowledge to guide the synthesis of training data, as in physics-guided generative adversarial networks that produce physically plausible synthetic samples.

The descriptive statistics assembled by the authors chart the field’s explosive growth. Publication output was modest between 2014 and 2016, began climbing steadily from 2017, and accelerated sharply after 2020, peaking at 78 journal articles in 2024. Geographically, China dominates with 165 of the reviewed publications, followed by the United States with 63, then the United Kingdom, Germany and India. Computer science accounts for 40 percent of the work and engineering 21 percent, but contributions also flow from materials science, mathematics, earth and planetary sciences, physics, medicine and molecular biology, underscoring the field’s inherently interdisciplinary character. A keyword co-occurrence analysis placed deep learning at the center of the research landscape, orbited by object detection, knowledge transfer, image segmentation and neural networks.

The practical payoff, the authors argue, is a rare combination of accuracy, interpretability and robustness that pure data-driven methods struggle to match. KGML reduces the volume of annotated data required, speeds convergence, strengthens resistance to adversarial manipulation, and produces predictions that remain consistent with established domain expertise even in out-of-distribution situations. The paradigm also dovetails with the broader goals of information fusion, since combining heterogeneous knowledge sources with visual data is fundamentally a knowledge-data fusion problem, playing out at the data, feature and decision levels.

Yet the survey is candid about the obstacles ahead. Representing heterogeneous knowledge, whether symbolic rules, ontologies, physical equations or an experienced radiologist’s tacit intuition, in a form neural networks can consume remains unsolved. Domain knowledge evolves, demanding mechanisms for dynamic updating without catastrophic forgetting. Computational complexity threatens scalability in time-sensitive applications such as video surveillance. Standard metrics like accuracy and mean average precision fail to capture whether a model is genuinely consistent with the knowledge it was given, prompting calls for new evaluation measures of knowledge consistency, semantic alignment and robustness under distribution shift. There is also a cautionary note on bias: knowledge itself can be prejudiced, and baking flawed domain assumptions into models risks entrenching inequities, particularly in medicine. The authors further acknowledge limitations of their own review, including reliance on Scopus and the exclusion of the most recent 2025 conference papers.

Looking forward, the authors see KGML as foundational to next-generation vision systems in medical imaging, autonomous driving, remote sensing and robotics, where predictions must be not only accurate but explainable and consistent with how the world actually works. They call for cross-dataset validation, more diverse real-world data, closer collaboration between machine learning researchers and domain experts, and evaluation frameworks that reward reliability rather than benchmark scores alone. If those challenges can be met, knowledge-guided machine learning may finally deliver what data-only approaches promised but never quite delivered: machines that see the world the way informed humans do.