AI Glossary / Computer vision and multimodal AI / Advanced

N-jet

Definition

An N-jet is the collection of a function's derivatives at a point up to a specified order N. In computer vision, it usually describes local image structure through partial derivatives of an image, often after Gaussian smoothing at a chosen scale. It provides a compact differential description of how image intensity behaves near that location.

What it means

The value of an image at one pixel says little about nearby structure. First derivatives describe local rates of change, while second derivatives describe how those rates change. Higher derivatives add further local information.

The letter N refers to the highest derivative order included, not the number of pixels in a patch. For a smooth two-dimensional scalar image, a full second-order jet includes the image value, two first derivatives, and three distinct second derivatives.

In practical image analysis, smoothing scale matters because a derivative estimated from tiny pixel variations can be dominated by noise. A multiscale jet records local behavior at different levels of spatial detail rather than treating one neighborhood size as universally correct.

How it works

Represent the image as a sufficiently smooth function L(x,y), commonly obtained by smoothing the sampled image. Choose the maximum derivative order and the spatial scale or scales of interest.

For a second-order jet, calculate L, Lx, Ly, Lxx, Lxy, and Lyy at each selected location. The mixed derivatives coincide under the usual smoothness assumptions, so Lxy and Lyx are not separate independent entries in this setting.

Use the derivative collection directly as features or combine its components into quantities such as gradient magnitude, Laplacian, or other differential descriptors. Some combinations have useful transformation properties, but the raw derivative vector is generally coordinate-dependent.

Compare representations under consistent scale, coordinate, and normalization conventions. Increasing derivative order increases information and numerical demands. It does not automatically make a descriptor more reliable, particularly when the image is noisy, coarsely sampled, or poorly matched to the smoothing scale.

Three practical examples

1. Local structure analysis: First and second derivatives help distinguish a locally flat region from a strong intensity transition or curved intensity pattern.

2. Image description: A vision pipeline builds feature vectors or histograms from derivative responses for recognition experiments.

3. Multiscale inspection: An analyst compares local derivative structure at fine and coarse scales to separate small surface texture from broader intensity variation.

Think of it this way

Think of describing a small stretch of road by its height, slope, and how quickly the slope changes. Those numbers summarize the local profile without listing every nearby elevation.

The analogy is one-dimensional, whereas images usually vary in two spatial directions. A jet also describes a local approximation rather than a complete map: distant structures and changes beyond the selected derivative order are not fully represented.

Worked example

Consider the smooth teaching function L(x,y) = 10 + 2x + 3y + x² + 4xy + 2y². At the origin, its value is 10, its first derivatives are Lx = 2 and Ly = 3, and its second derivatives are Lxx = 2, Lxy = 4, and Lyy = 4.

Its second-order jet at that point is therefore (10,2,3,2,4,4), using the stated ordering. The associated Taylor expression is 10 + 2x + 3y + one-half of (2x² + 8xy + 4y²).

At (0.1,0), this gives 10.21. Because the teaching function is exactly quadratic, the second-order expression is exact here. For a general image function it would only be a local approximation, with higher-order terms omitted.

The example also shows that the mixed derivative contributes to the cross term with the correct Taylor coefficient; it is not simply another independent pixel measurement.

Illustrative case study

Fictional case: A materials researcher studies texture differences in microscope images of manufactured coatings. A first derivative-based descriptor responds strongly to image noise at the smallest scale.

The researcher computes derivative collections after several controlled smoothing levels and evaluates them on held-out specimens. The report identifies the scale and normalization used for each experiment, rather than treating “N-jet features” as a complete specification.

Some higher-order features add little useful discrimination under the available resolution. The team retains the simpler representation where appropriate and checks whether apparent differences arise from coating structure or inconsistent image acquisition. No recognition improvement is assumed merely from adding derivative channels.

Common misconceptions and mistakes

Do not interpret N as patch width or feature-vector length. In two dimensions, the number of derivatives through order N grows according to the combinations of horizontal and vertical derivative orders.

Do not assume a jet is automatically invariant to rotation or scale. Such behavior depends on normalization, coordinate handling, and how components are combined.

Do not differentiate noisy data without considering smoothing and discretization. Higher-order derivatives can be particularly sensitive to those choices.

Do not confuse a local approximation with global image reconstruction. A finite jet at one point generally leaves many possible surrounding images consistent with the same derivative values.

Business use

N-jets are mainly a technical foundation for visual descriptors and differential image analysis. They give developers an interpretable vocabulary for local structure and can support compact feature construction when a task benefits from explicitly controlled spatial scales.

For a business application, the practical question is whether the resulting representation improves the decision under realistic acquisition conditions. Compare accuracy, robustness, computational cost, and ease of calibration. Record image resolution and scale conventions so a model trained on one camera setup is not silently given a different physical neighborhood on another.

People also ask

Does a second-order jet include first derivatives?

Yes. A full jet through second order includes lower orders as well.

Are N-jets exclusively neural-network features?

No. They are mathematical local descriptions used in many kinds of image analysis.

Does more smoothing always help?

No. It can reduce noise while also removing small structures needed for the task.

Check your understanding

How many components are in the full second-order jet of a smooth scalar two-dimensional image, counting its value?

Answer: Six: the value, two first derivatives, and three distinct second derivatives.

Related terms

• Scale space
• Taylor expansion
• Image gradient
• Hessian
• Gaussian derivative
• differential invariant
• local image descriptor

Learn next

Differentiate a simple quadratic surface and assemble its second-order jet. Then estimate derivatives on a small image at two smoothing scales and inspect how noise, edges, and broader structures change.

More in Computer vision and multimodal AI

• Multimodal instruction-tuned - Multimodal instruction-tuned describes a model trained to respond to instructions while processing inputs beyond text, such as images, audio, or video.
• Multimodal model - A multimodal model is a machine-learning model whose inputs, outputs, or both involve more than one type of information, such as text, images, audio, or video.
• Multiplicative noise - Multiplicative noise is random variation that scales a signal rather than simply adding a fixed-size disturbance to it.
• Naka - Rushton curve - The Naka - Rushton curve is a saturating response function used to describe how a response grows with stimulus intensity and approaches a maximum.
• NAND operator - The NAND operator means NOT AND. For two Boolean inputs A and B, it returns false only when both inputs are true; otherwise it returns true.
• Natural image statistics - Natural image statistics are measurable regularities in images of real-world scenes.

Sources

Linde and Lindeberg, Composed Complex-Cue Histograms (2012), author-hosted abstract and highlights on Gaussian derivatives and N-jet descriptors: https://www.csc.kth.se/~tony/abstracts/LinLin11-CompComplCueHist.html . The primary author page was read; the numerical Taylor example is original and is not a reproduction of its experiments.

Previous: n-gram model Next: N-of-1 trial