Image processing turns an image into another image, enhanced or filtered, while computer vision turns it into an analysis of its content. The two are not rival techniques. They are layers of one pipeline, and in art the top layer produces probabilistic evidence that sits beside an expert’s judgement rather than a verdict on its own.
Computer vision vs image processing, seen on a painting
Image processing works on the picture without assuming anything about what it shows: contrast enhancement, edge extraction, noise removal and geometric transforms change the image but produce no interpretation of its content. The OpenCV documentation describes its imgproc module as a collection of per-pixel operations, filters and geometry transformations useful for computer vision. David Stork’s survey of computer analysis of paintings draws the same line, calling point-based colour and brightness adjustment “better described as image processing than as image analysis”.
Computer vision tries to describe the world an image depicts and reconstruct its properties. Szeliski frames it as an inverse problem: the image underdetermines the scene, so a system has to fall back on physics-based or probabilistic models, or on machine learning from many examples, to choose between possible explanations.
Gonzalez and Woods treat the whole field as a continuum: low-level processes have images as both input and output, while mid-level processes take images in and return attributes such as edges, contours or the identity of objects. The boundary between the layers keeps moving. Deep learning models now do denoising and super-resolution, so one network can handle cleanup and interpretation, although the order stays the same: the input has to be stable before the system can trust what it reads.
| Image processing | Computer vision | |
|---|---|---|
| Output | An image: enhanced, filtered, transformed | An analysis of content, or a decision based on it |
| Assumes anything about the content? | No | Yes, through a model of the scene or learned examples |
| Painting examples | Canny, Sobel, Laplacian and Scharr edge maps; grayscale conversion; colour adjustment; area filtering to show properties the unaided eye misses | Style and movement classification; verifying a work against a target artist; locating zones painted by a master or workshop |
| Works best when | Input is consistent and the task fits clear mathematical rules | The task involves interpretation, ambiguity or changing conditions |
| Can stand alone? | Yes, when a person reads the result (an enhanced infrared image for a conservator, say) | It depends on clean, consistent input from the cleanup layer below it |
If you want the general mechanics rather than the art case, our walkthrough of how computer vision works step by step covers the stages one at a time.
What can computer vision actually detect or measure in a painting?
What a model reads in a painting is visual input such as shapes, edges and colours, not sociocultural context. A 2026 arXiv study on authenticating historical drawings built its feature vector from Fourier-domain energy, Shannon entropy, global contrast, grey-level co-occurrence (GLCM) homogeneity and a box-counting estimate of fractal complexity. Each of those features reads a property of the drawing: Fourier energy separates fine mark-making from broad tonal masses, entropy tracks how varied the tonal values are, and fractal dimension rises towards 2 for dense crosshatching and falls towards 1 for simple line work.
Brushstrokes are highly informative for attribution, and they often show up in an image as edges. Machine learning can also pick up micro-level features of painterly technique that the human eye does not see, which is why a 2025 study of Rubens’ workshop trained on those details deliberately.
Beyond texture, Stork’s survey records computer methods for perspective, lighting and three-dimensional reconstruction, including a 3D model of Georges de la Tour’s Christ in the carpenter’s studio that placed the light source at the candle and argued against the claim that the painting was made by optical projection. Craquelure, the network of fine cracks on a paint surface, has also been characterised by image analysis for retrieval.
How is computer vision used in painting and the visual arts?
Computer vision in the visual arts does four jobs: it classifies style, it supports attribution, it maps where different hands worked on one canvas, and it makes large collections searchable.
Classifying style and movement
Style classification treats artistic movement as a label a network learns from labelled examples. Lecoutre and colleagues used deep residual networks on the Wikipaintings dataset and reached 62% accuracy across 25 styles, outperforming existing approaches. Karayev and colleagues built a dataset of 85,000 paintings labelled with 25 style and genre categories and found that features learned in a multi-layer network generally performed best, even when that network had been trained on object labels rather than style.
Movements blend into each other over time, one artist’s paintings can differ in style, and some artists belong to no single period, which is exactly why classification is hard in this domain. Another style study made a related point from its confusion matrix: misclassifications fell between historically related styles, which the authors read as evidence the CNN had learned relevant features.
Searching museum collections
The Museum of Modern Art’s Digital Media team worked with Google Arts & Culture Lab, using machine learning, to get a head start on identifying artworks in tens of thousands of exhibition photographs going back to 1929, so that a 1929 installation shot now links to the Cézanne it shows. Lecoutre and colleagues note that correctly identifying a painting’s style also matters for indexing large art databases.
Why models trained on photographs need retraining for paintings
In transfer learning, the early layers of a network capture general features such as colour variation, tonal values, edges and textures, and they can be frozen, while how many deeper layers need fine-tuning depends on how similar the source task is to the target task. A Raphael study in Heritage Science (2023) used exactly this approach: ResNet50, trained on ImageNet for object recognition, with its top layers removed and its feature vectors fed to a support vector machine (SVM).
The gap shows up in specifics. In the 2025 Veronese study published in Frontiers in Artificial Intelligence, a general-purpose MobileNetV2 had to upsample 64×64 patches to 224×224 pixels, and that upsampling acted as a low-pass filter that smoothed away the high-frequency microtexture that distinguishes Veronese’s brushwork. Restricted to RGB input, the same model also lost the grayscale and edge-map channels that carried decisive information about brushstroke dynamics. The authors concluded that for specific heritage tasks a tailored, texture-focused model was more resource-efficient and more reliable than general-purpose large models.
That matches what we hold about production computer vision in general: off-the-shelf models that work in demos fail at the edge cases and conditions of real deployments. The decision between hand-crafted features and learned ones is covered properly in our piece on computer vision and pattern recognition; for paintings, hand-crafted features keep advantages when data is scarce.
Can AI tell whether a painting is real or fake?
AI does not settle whether a painting is real or fake; it gives probabilistic evidence of how closely a painting’s style and execution match works accepted as one artist’s. The Raphael and Veronese systems are layered, with image processing producing edge features and a learned model reading them.
Deep features from ResNet50 alone sometimes failed to tell imitations from authentic works, so the authors added a second check on combined Canny, Sobel, Laplacian and Scharr edge features, and a painting counted as likely Raphael only if it passed both the edge threshold and the SVM probability threshold. On a disputed painting the model returned a probability of 0.57 that the whole work was Raphael’s, which the authors read as supporting scholars who argue for shared attribution with Giulio Romano.
The Veronese study fed 64×64 patches with five channels (RGB, grayscale and a Canny edge map) to a shallow network, which turned a handful of paintings into 4,374 training patches. Cross-validated accuracy was 94.51%, authentic Veronese works scored between 80.3% and 99.9% at painting level, and non-Veronese works scored between 11.9% and 36.7%. A copy that matched an authentic painting’s composition scored 11.9% against the original’s 89.2%, so the model separated stylistic coherence from compositional similarity. The painting actually under dispute scored 61%, and the authors said plainly that this was not enough for a definitive attribution.
What the authentication numbers license
The 2026 drawings study reported a pooled true acceptance rate of 83.3% and a false acceptance rate of 9.5% at a threshold chosen to keep false accepts low, because accepting a non-target work costs more than flagging a genuine one for expert review. Its false accepts clustered in a few confusable pairs of artists, which reflects shared training lineages and drawing conventions as much as any algorithmic weakness.
Each of these studies limits its own claim. A full authentication protocol relies on provenance, material studies, iconography and condition, and the Raphael authors describe their work as covering only the connoisseurship part. The Rubens authors warn of erroneous conclusions in cases such as an authentic signature on a pupil’s work or an especially skilful forgery. General computer vision also treats “flag for further human review” as a standard decision outcome.
What computer-vision techniques are most used in art-conservation labs?
In conservation labs, spectral imaging is a non-invasive tool whose output is processed images and material maps that a conservator reads, rather than an automated interpretation. A review of spectral imaging in conservation names infrared reflectography as the technique most widely used to image and study underdrawings. Multispectral and hyperspectral imaging record spectral and spatial information together, so materials can be mapped across the surface instead of sampled at single points.
The processing layer here groups pixels with similar spectral signatures and identifies materials from them. It has revealed things the eye cannot: in one painting it separated white pigments clearly enough to show a later repainting that was invisible without it. Per-pixel spectral data also allows virtual restoration of original colours and simulated varnish removal.
The limits are just as concrete. Results for identifying pigments and their mixtures were mixed, mainly because of restricted spectral range and the lack of a comprehensive database of pigment reflectance spectra, so invasive micro-sampling is still needed to characterise individual materials and paint layers. Choosing equipment depends on the job: a high-resolution camera matters most for documentation, while spectral range and number of bands matter most for pigment identification.
Why paintings are a hard domain for computer vision
Paintings are hard for computer vision partly because style labels rest on expert judgement. Style is assigned by historians from subjective attributes, physical characteristics and historical period. Expert visual analysis is also exposed to intuition, biased expectations and institutional narratives.
The model observes less than the expert does. A CNN predicts from shapes, edges and colours with no knowledge of sociocultural context. The Rubens model was trained only on 512×512 tiles and never saw a whole painting, so high image-level accuracy did not mean it had learned composition. The Veronese authors note the same limit for patch-based learning: no overall composition, figure-scale relationships or large-scale colour organisation.
Data is scarce in historical attribution, and a single plausible measure can still be contested. Hand-crafted features keep real advantages when samples are few, because they need less data, can be inspected by experts and hold up better under distribution shift. Classical analytical methods are faster, cheaper and easier to maintain than learned models when the task fits clear rules, and one review concluded that traditional computer vision is not obsolete and that hybrid methods help on problems deep learning does not suit. Pollock’s drip paintings show the risk of leaning on one measure: Taylor and colleagues proposed that they have fractal structure, and several scholars have questioned whether simple fractal dimension can tell genuine Pollocks from forgeries.
Does generative AI threaten the painting market?
The painting-analysis studies do not measure the art market. What they do show is narrower. In the Veronese study, a copy faithful to the original’s composition still scored low, at 11.9%. The Rubens authors still list a skilful forgery that carefully imitates the original as a case where machine analysis can mislead.
Building a layered analysis pipeline for a collection
Between them, these studies use deterministic edge operators, a shallow model sized to a small training set, a threshold set where false accepts cost more, and visualisations such as probability heatmaps that show how a decision was reached. Where a museum, archive or conservation lab wants that kind of pipeline built and tested on its own image sets, TechnoLynx’s Computer Vision R&D team does that engineering work. We would claim the engineering, not the attribution: the art-historical judgement stays with the specialists who own it. For the step from classical operators to learned models in general, our overview of AI and computer vision covers the model families.
The model’s score is a reason to look again
The Veronese study states the conclusion plainly: the network supplies strong probabilistic evidence of stylistic affinity, and conclusive attribution needs technical, historical and provenance analysis beyond what computation can infer. The Rubens authors put the tool’s role as raising questions and prompting closer examination, especially where surface evidence is inconclusive. That sets the standard for building one of these systems. Measure it by whether it shows an expert where to look and why (a heatmap over a Joseph’s face, a flagged zone of workshop hands), not by whether its accuracy figure looks like a verdict.
FAQ
What does computer vision include?
Computer vision fundamentals cover acquiring, processing, analysing and understanding digital images in order to produce numerical or symbolic information, such as decisions. Image processing is one of those stages, not a separate field that competes with it.
Is image recognition the same as image processing?
No. Image recognition labels what an image contains, while image processing changes pixels and labels nothing. Recognition algorithms themselves are covered in our guide to image recognition.
