AI Learns to Inspect the World With Far Fewer Labels

A new review of more than 170 peer-reviewed studies suggests that artificial-intelligence systems could become far more practical for visual inspection by learning from images that human experts have never labelled. The approach, built around semi-supervised and self-supervised learning, is attracting attention across medical imaging, industrial maintenance and airport security because it addresses one of the most stubborn bottlenecks in modern computer vision: the enormous cost of telling an algorithm exactly what it is looking at. The review, published in the International Journal of Machine Learning and Cybernetics, finds that label-efficient systems can in many cases approach the performance of fully supervised models while requiring only a fraction of the painstaking annotations traditionally used to train them.

Visual inspection sounds simple when performed by a person. A radiologist may identify a suspicious region in a scan, an engineer may spot a crack in a turbine component, or a security officer may notice a prohibited object in an X-ray image. For an algorithm, however, every example must be converted into structured information. A training image might need a category label, a pixel-by-pixel segmentation mask, or a bounding box showing the exact position of a defect. Producing those annotations requires time, technical expertise and, in high-stakes settings, professional verification. The problem becomes especially severe when defects are rare, visual conditions change or the objects of interest are partly hidden by clutter and overlapping materials.

Semi-supervised learning attempts to reduce that burden by combining a relatively small collection of labelled images with a much larger pool of unlabelled examples. The labelled data provide an explicit learning signal: they tell the model which visual patterns correspond to a tumour, fracture, threat object or equipment fault. The unlabelled data are then used to refine the model’s understanding of the underlying image distribution. In practice, a system may first make predictions for unlabelled images and then use only sufficiently reliable predictions, known as pseudo-labels, during further training. Other methods encourage the model to produce consistent answers when the same image is altered through cropping, rotation, colour changes or other augmentations.

Self-supervised learning goes a step further by creating a learning task from the data themselves. Instead of asking humans to annotate every image, researchers design a pretext problem whose answer can be inferred automatically. A neural network might be trained to reconstruct a masked section of an image, determine whether two altered views came from the same original, predict the temporal order of frames or distinguish genuine visual structure from deliberately corrupted input. These tasks force the model to learn representations—numerical patterns in a high-dimensional feature space—that capture edges, textures, shapes and spatial relationships. After this pretraining stage, the representation can be adapted to a specific inspection task with comparatively few labelled examples.

The review by Saad Mazhar Khan and colleagues maps how these ideas have moved from general computer vision into domains where mistakes can carry serious consequences. In medical imaging, self-supervised representations can help models work with small collections of annotated scans, including images in which abnormal tissue occupies only a tiny portion of the field of view. In industrial inspection, systems can learn what normal surfaces, components or production patterns look like and flag deviations as anomalies. In baggage screening, semi-supervised detectors must contend with dense, overlapping objects in X-ray imagery, where a threat may be heavily occluded and its appearance may vary with orientation, material and scanning conditions.

Anomaly detection is particularly well suited to label-efficient learning because the definition of “abnormal” is often difficult to enumerate in advance. A factory may produce millions of normal components but encounter only a handful of examples of a newly emerging fault. Rather than training a classifier on every possible defect, an algorithm can learn a compact representation of normal data and calculate how far a new image lies from that learned pattern. Autoencoders, for example, compress an image into a latent representation and attempt to reconstruct it; unusually large reconstruction errors may indicate an anomaly. Other systems compare deep feature embeddings, using distances in representation space to identify images or regions that do not resemble the normal examples.

The review also highlights the growing role of contrastive learning, masked autoencoders and transformer architectures. Contrastive methods bring representations of related images closer together while pushing unrelated examples apart, allowing a model to learn which changes preserve an object’s identity. Masked autoencoders hide portions of an image and train the network to infer the missing content, encouraging it to capture both local texture and broader context. Vision transformers divide images into patches and use attention mechanisms to estimate which regions are related, potentially allowing them to recognize a defect even when it is separated from the most obvious visual cues. These methods are increasingly being combined with pseudo-labelling, knowledge distillation and domain-adaptation strategies.

The central challenge is that an inspection system rarely encounters exactly the same visual world as the one used during training. A medical scanner may differ between hospitals, an industrial camera may be replaced, lighting can change on a production line and airport X-ray systems may produce different colour mappings or resolutions. This phenomenon, known as domain shift, can cause a model to mistake changes in equipment or environment for meaningful changes in the object itself. Cross-domain methods try to align the feature distributions of a source domain and a target domain, sometimes without access to target labels. The goal is to preserve task-relevant information while making the representation less sensitive to irrelevant differences in acquisition conditions.

That transfer is not guaranteed, and the review stresses several unresolved weaknesses. Pseudo-labels can amplify early mistakes: if a model confidently assigns the wrong label to an unlabelled image, subsequent training may reinforce the error. Rare defects and severe class imbalance can make the system appear accurate while missing the cases that matter most. Occlusion and clutter can hide crucial evidence, while adversarial inputs or poisoned training data may deliberately manipulate a model’s predictions. Interpretability remains another obstacle. A heat map showing where a model focused is not necessarily an explanation of why it made a decision, and a system that cannot communicate its uncertainty may be difficult to trust in clinical or safety-critical environments.

Security and privacy become more complicated when inspection data are shared across institutions. Medical images can contain sensitive information, while industrial and airport datasets may be commercially or operationally confidential. Federated learning offers one possible route: separate organisations train a shared model locally and exchange model updates rather than raw images. Yet the review notes that model updates can themselves leak information, and malicious participants may introduce backdoors or corrupt the learning process. Secure aggregation, robust validation and careful monitoring are therefore necessary if label-efficient systems are to move from research demonstrations into dependable operational tools.

The authors identify multimodal learning as one of the field’s most promising directions. Future inspection systems could combine images with text, sensor readings, equipment histories, three-dimensional geometry or other forms of contextual information. A defect that is visually ambiguous might become easier to classify when paired with vibration data from a machine or the operating conditions under which an image was captured. Transformer-based models could provide a common architecture for fusing these signals, while self-supervised objectives could exploit relationships among modalities without requiring every example to be manually annotated. Such systems may also support more informative explanations by connecting a visual decision with comparable cases, physical measurements or known failure modes.

The review’s broader message is not that human annotation is becoming unnecessary, but that it can be used more strategically. Experts could concentrate on uncertain or high-impact cases selected by an active-learning system, rather than labelling images at random. Human reviewers could correct pseudo-labels, define safety thresholds and test models against deliberately difficult examples. Evaluation would also need to extend beyond headline accuracy to include sensitivity to rare events, calibration of confidence, robustness under domain shift, computational efficiency and performance across different populations or operating environments.

Because the study is a systematic review rather than a new experimental benchmark, it does not introduce a dataset or report a single winning algorithm. Its contribution is to draw together evidence from more than 170 studies published from 2016 onward and compare the recurring patterns across different inspection domains. The synthesis shows why semi-supervised and self-supervised learning have become so influential: they offer a way to turn the vast quantity of unlabelled visual data already collected by hospitals, factories and security systems into useful training material. But it also makes clear that lower annotation costs do not automatically produce safer or fairer systems.

For visual inspection, the next breakthrough may therefore come less from teaching machines to recognize one more category and more from teaching them how to learn responsibly from imperfect evidence. Models that can discover structure without labels, adapt across environments and reveal when they are uncertain could make automated inspection faster and more accessible. Models that silently absorb bias, magnify pseudo-label errors or fail under unfamiliar conditions could do the opposite. The emerging technology promises a dramatic reduction in the labour of annotation, but the review argues that rigorous validation, human oversight and transparent evaluation will determine whether that promise becomes a reliable tool—or merely another impressive laboratory demonstration.