Filter Visualisation Calculator

Visualisation results

Results update after processing the current image and filter configuration.

Ready

Load an image or sample, configure a filter, then select Process Filter.

—Input dimensions

—Output dimensions

—Mean activation

—Maximum response

—Output sparsity

—Multiply-adds

Original image

Filtered output

Activated output

Pooled output

Absolute response

Receptive field

Click a filtered output cell to inspect its source patch.

Before-and-after comparison

No processed results yet.

Top responding locations

No responses available.

Output matrix preview

No matrix available.

Select an output position by clicking the filtered image.

Kernel position	Input value	Kernel value	Product
No selected calculation.
Total before bias/divisor			—

Add filters from the configuration panel.

1. Image and input data

Upload image

PNG, JPG, and WebP files are supported.

Built-in sample

Input channel

Working width

Working height

Display interpolation

Custom image matrix

2. Filter kernel

Filter preset

Kernel size

Divisor

Bias

Filter scale

Editable coefficients

Learned-filter import

Filter bank name

3. Convolution, activation, and pooling

Output mode

Operation

Padding mode

Stride X

Stride Y

Decimals

Padding X

Padding Y

Dilation X

Dilation Y

Clip minimum

Clip maximum

Activation

Activation parameter

Pooling

Pool W

Pool H

Pool stride

Round numerical output

4. Visualisation and comparison controls

Heatmap colour scale

Canvas scaling

Display threshold

Brightness

Contrast

Gamma

Top patches

Comparison filter

Enable filter comparison

Reverse colour scale

Show heatmap values

Invert display

Formula used

The output size is calculated independently for width and height.

Output = floor((N + 2P − D(K − 1) − 1) / S) + 1

Here, N is input size, P is padding, D is dilation, K is kernel size, and S is stride.

How to use

Upload an image, choose a sample, or paste a numerical matrix.
Select a preset filter or edit every kernel coefficient manually.
Configure stride, padding, dilation, activation, pooling, and display controls.
Select Process Filter, then inspect images, heatmaps, statistics, and calculations.
Export images, matrices, filter weights, configurations, or a printable report.

Example configurations

Goal	Kernel	Padding	Activation	Expected result
Vertical edges	Sobel X	Same, reflection	Absolute value	Strong vertical boundaries
Image smoothing	Gaussian blur	Same, replicate	None	Reduced high-frequency texture
CNN-like activation	Random 3×3	Same, zero	ReLU	Sparse positive feature map
Downsampled features	Sharpen	Same, zero	ReLU plus max pooling	Compact salient responses

Filter visualisation concepts

A convolution kernel scans local image patches and calculates weighted sums. Learned CNN filters often respond to edges, textures, colours, shapes, or task-specific patterns.

Stride controls movement between patches. Padding controls border treatment, while dilation expands spacing between kernel elements.

Feature maps show where a filter responds strongly. Activation functions transform responses, while pooling compresses local information.

Frequently asked questions

What is a filter kernel?

It is a small matrix of weights applied repeatedly across an image or feature map.

Why can feature-map values be negative?

Negative values indicate opposite directional or pattern responses relative to positive values.

What is cross-correlation?

Cross-correlation applies kernel weights directly. Mathematical convolution flips the kernel first.

What does same padding mean?

Same padding chooses border padding intended to preserve spatial size when stride equals one.

What does dilation change?

Dilation spreads kernel samples apart, increasing the receptive field without increasing parameter count.

Why use ReLU?

ReLU removes negative responses and often produces sparse, easy-to-interpret activation maps.

How is sparsity calculated?

The calculator reports the percentage of values whose magnitude is effectively zero.

Can this calculator load learned CNN filters?

Yes. Paste a two-dimensional JSON array or CSV matrix into the learned-filter field.

What is a receptive field?

It is the source image region contributing to one selected output position.