MuhammadLab
Deep learning visual tool

Image Convolution Interactive Tool

See how a 3x3 convolution kernel slides over pixels, multiplies neighboring values, and creates a feature map. This is the same core idea behind early CNN layers for edges, textures, shapes, and visual patterns.

CNN kernelsPixel mathLocal image upload

Active kernel

Edge Detection

-1
-1
-1
-1
8
-1
-1
-1
-1

Output pixel = sum of neighboring pixels multiplied by the kernel weights.

Image mode

Apply the kernel to an image

Use the generated teaching image or upload a PNG, JPG, or WebP. Processing stays local in your browser.

Current image: Generated teaching image

Original image
Convolved output

Pixel-grid mode

Watch the convolution calculation

Input patch

20
25
30
34
38
42
45
22
28
36
48
58
62
66
25
36
80
128
150
76
70
30
52
125
220
200
88
75
28
44
96
170
160
82
72
24
32
52
72
76
68
64
20
24
30
38
44
50
55

Kernel weights

-1
-1
-1
-1
8
-1
-1
-1
-1

Center pixel example

Raw sum: 651.00

Clamped output: 255

Output feature map

85
64
69
62
60
67
190
42
0
0
0
0
0
233
32
0
0
107
255
0
193
55
0
174
255
255
0
212
42
0
1
255
255
0
199
44
0
0
0
0
0
185
80
34
22
30
48
93
255

Center-pixel multiplication

80 x -1 = -80.00
128 x -1 = -128.00
150 x -1 = -150.00
125 x -1 = -125.00
220 x 8 = 1760.00
200 x -1 = -200.00
96 x -1 = -96.00
170 x -1 = -170.00
160 x -1 = -160.00

CNN intuition

Convolution turns local pixel neighborhoods into feature maps.

In a neural network, kernels are learned from data instead of manually selected. Early layers often respond to edges and textures; deeper layers combine those responses into shapes, objects, and task-specific patterns.

Continue to CNNs

Teaching note

This tool uses hand-picked kernels for learning. Real CNN filters are initialized and then learned through backpropagation, so the network discovers useful patterns from training examples rather than relying on fixed filters.