👁️ Computer Vision: CNN Pipeline

About this simulation

A convolutional neural network "sees" an image by sliding a small filter, or kernel, across the pixel grid. At each position it multiplies the kernel weights against the pixels underneath and sums the result — a single number that becomes one cell of the feature map. Stacking many filters and pooling layers lets a network build up from raw pixels to edges, then shapes, then whole objects.

How it works

• The front plane is the raw input image — a synthetic bright blob on a noisy background, standing in for an object to detect.
• A glowing wireframe box is the kernel window; it scans left-to-right, top-to-bottom exactly like a real convolution.
• Each scanned position lights up one cell of the feature map plane behind it, colored by the convolution response (after ReLU).
• The pooled map further back applies 2×2 max pooling, keeping only the strongest activation in each block — this is how CNNs shrink spatial size while keeping the strongest signals.
• The detection box projects the pooled map's strongest region back onto the input image, mimicking how object detectors localize a target.

Controls

• Kernel size — 3×3, 5×5 or 7×7 receptive field for each convolution step.
• Stride — how many pixels the kernel jumps between steps; larger strides give a coarser feature map.
• Image noise — sensor-style noise added to every pixel before convolution.
• Scan speed — how fast the kernel sweeps across the image.
• Filter kernel — swap between an edge-detecting, blurring, or sharpening filter to see how the same image yields different feature maps.
• Detection box — toggle the bounding box drawn around the highest-activation region.

Did you know?

The first influential CNN, LeNet-5 (LeCun et al., 1998), used exactly this convolution → pooling pattern to read handwritten digits on bank cheques — the same basic recipe still powers modern object detectors and image classifiers today.