AI Glossary / Computer vision and multimodal AI / Advanced

Convolution

Definition

Convolution combines two functions or sequences by summing or integrating products at relative shifts. For discrete one-dimensional sequences, y[n]=Σk x[k]h[n−k]. In image processing, a fixed convolution kernel can produce each output value from a weighted neighborhood of input values. The same mathematical idea also appears in signals, probability and neural networks.

What it means

Convolution provides a consistent rule for combining nearby information. The chosen weights determine what the rule does: they might average neighboring values or respond to changes. The operation itself is not synonymous with blurring and is not necessarily learned. A designer can choose fixed coefficients, while a neural-network training process can learn weights. Understanding the weights, input convention and boundary handling is essential before interpreting the output. A familiar operation name alone does not tell you what a particular processing stage will preserve or suppress.

How it works

For each output index, align the sequences according to the convolution formula, multiply overlapping values and sum their products. In multiple dimensions, the summation extends over the relevant coordinates. Finite arrays require a convention for values outside their boundaries and a choice of output region. True convolution includes reversal of the kernel's coordinate direction relative to cross-correlation. Many neural-network layers use cross-correlation under the name convolution, so check the actual definition when reproducing a calculation or transferring a fixed asymmetric kernel between tools.

Three practical examples

• An image-processing team uses a normalized averaging kernel to reduce small fluctuations before inspecting the effect on fine detail.
• A sensor-analysis team combines a sampled signal with a specified filter response, checking timing and boundary conventions before interpreting the result.
• A developer compares a hand calculation with a library operation to confirm that kernel orientation and output size match the intended processing rule.

Think of it this way

Imagine a reusable scoring stencil that combines several neighboring numbers into one result. Sliding it creates a new list of results, each based on the same rule. The exact orientation of the stencil matters when its weights differ from one side to the other. This analogy helps explain repeated weighted combinations, but the convolution formula is the authority for alignment. It also leaves room for different stencils that emphasize different kinds of information.

Worked example

In an original finite-sequence example, let x=[2,4,1] and h=[1,2], with both sequences beginning at index zero and equal to zero elsewhere. Full convolution gives y[0]=2×1=2; y[1]=2×2+4×1=8; y[2]=4×2+1×1=9; and y[3]=1×2=2. Thus y=[2,8,9,2], with length 3+2−1=4. These weights are not an average because they sum to three. Replacing h by [1/3,2/3] scales every result by one third. The example demonstrates alignment, output length and weight scaling; it does not describe a complete image filter or guarantee that either kernel suits a particular signal.

Illustrative case study

A fictional imaging team ports a processing step to a different library. Most test pictures look similar, but a directional edge response changes sign. The team compares a tiny asymmetric kernel on a small numerical array and discovers that one interface applies correlation while the intended calculation uses convolution. After resolving orientation, it examines borders and output types separately. The team keeps this small example as a reproducibility check for the pipeline. The case illustrates why matching an operation's name is insufficient; it does not claim that either library is incorrect or that the sign difference would matter equally in every application.

Common misconceptions and mistakes

Convolution is not ordinary matrix multiplication, though it can be represented by a specially structured matrix. It is not always a smoothing operation, and its weights need not sum to one. Another error is overlooking padding: zero, reflected and wrapped boundaries encode different assumptions. Output types also matter when results can be negative or exceed an image's storage range. Finally, a convolutional layer may include bias, channel mixing, stride and nonlinear activation elsewhere in the network; those details should not be silently folded into the bare mathematical definition.

Business use

Convolution is a foundational operation behind many image, audio and sensor-processing systems. Business users benefit from knowing which effect a configured or learned filter is intended to achieve and how that effect is evaluated. For a production pipeline, document conventions that affect reproducibility and assess the downstream task rather than an intermediate picture alone. Runtime depends on data sizes and implementation choices, so compare realistic workloads. When replacing software, verify numerical behavior and operational outcomes before assuming that two similarly named functions are interchangeable.

People also ask

Does convolution always blur an image?

No. Its effect depends on the kernel; some kernels emphasize differences instead.

Is cross-correlation identical to convolution?

Their alignment conventions differ, although some symmetric kernels produce the same result.

Must the output have the input's size?

No. Full, same-sized and valid-region conventions return different portions of the result.

Check your understanding

Why did the example produce four values?

Answer: Full convolution of lengths three and two has length four.

Why was the kernel not an average?

Answer: Its weights summed to three rather than one.

What should be checked when a directional response changes sign across tools?

Answer: Kernel orientation and whether the operation is convolution or correlation.

Related terms

• Convolution kernel
• Convolutional neural network
• Fourier transform
• cross-correlation
• padding
• linear filtering

Learn next

Calculate full convolution for two short sequences and then identify a same-sized output region under an explicit alignment convention. Repeat with a symmetric and an asymmetric kernel. Compare the corresponding correlation outputs before moving to two-dimensional images.

More in Computer vision and multimodal AI

• Contrast sensitivity function (CSF) - A contrast sensitivity function describes how sensitivity to visual contrast varies with properties such as spatial frequency under specified viewing...
• Contrast stretching - Contrast stretching remaps an image's intensity values so a selected input interval occupies a desired output interval.
• Convex hull - The convex hull of a set of points is the smallest convex set containing them.
• Convolution theorem
• Convolutional filter - A convolutional filter is a set of weights applied repeatedly across an input to produce responses to local patterns.
• Convolutional operation - A convolutional operation applies shared filter weights across an input to produce an organized set of local responses.

Sources

SciPy, Signal Processing tutorial, convolution and correlation: https://docs.scipy.org/doc/scipy-1.15.2/tutorial/signal.html OpenCV, Image Filtering documentation, filter2D convention: https://docs.opencv.org/4.11.0/d4/d86/group__imgproc__filter.html The numerical sequences, stencil analogy and software-porting scenario are original educational illustrations. Library details refer to the cited interfaces.

Previous: Convex set Next: Convolution kernel