CNNs and computer vision

In plain English

Convolutional Neural Networks (CNNs) treat images as grids of numbers (pixel intensities per color channel). Instead of connecting every pixel to every neuron, they reuse small filters scanned across the image.

That local connectivity matches how nearby pixels relate—edges, textures, parts—and drastically cuts parameter count compared to fully connected layers.

CNNs powered the jump in image classification, object detection, medical imaging, and later fed into multimodal systems that read images plus text.

How CNN layers stack

Convolution layers apply learnable filters, producing feature maps that highlight patterns. Stride and padding control output size. Pooling downsamples spatial resolution, adding rough translation tolerance.

Deep stacks alternate conv + nonlinearity + pooling, then global pooling or flattening feeds classifiers. Modern variants use depthwise separable convs or attention hybrids for efficiency.

Data augmentation (random crops, flips, color jitter) artificially expands training diversity and fights overfitting on vision tasks.

# 3x3 image and 2x2 filter — output is one number per valid position
image = [
    [1, 0, 1],
    [0, 1, 0],
    [1, 0, 1],
]
filt = [[1, 0], [0, -1]]

def conv2d(img, f):
    h, w = len(img), len(img[0])
    fh, fw = len(f), len(f[0])
    out = []
    for i in range(h - fh + 1):
        row = []
        for j in range(w - fw + 1):
            s = sum(img[i + di][j + dj] * f[di][dj]
                    for di in range(fh) for dj in range(fw))
            row.append(s)
        out.append(row)
    return out

print(conv2d(image, filt))

Going deeper

Detection and segmentation add bounding boxes, masks, and anchor-free heads—same conv backbone, different task heads.

Vision transformers patchify images and apply self-attention; hybrids combine conv inductive biases with global attention.

Responsible deployment considers demographic performance gaps, adversarial patches, and consent for facial recognition use cases.

Common misconceptions

CNNs understand objects like people do.

They learn statistical textures and shapes that correlate with labels; they can fail on unusual poses or backgrounds.

Vision transformers replaced CNNs entirely.

Both coexist; choice depends on data scale, latency, hardware, and task.

Key facts

• CNNs use shared local filters slid across spatial inputs.
• Parameter sharing and locality exploit image structure efficiently.
• Pooling and stride build hierarchical, translation-tolerant features.
• Data augmentation is standard practice in vision training pipelines.
• Modern vision stacks include CNN, transformer, and hybrid architectures.