Knowledge Base - OBSESSLABS | OBSESSLABS

The Vision Revolution

Convolutional Neural Networks (CNNs) have been the backbone of computer vision since AlexNet in 2012. By mimicking the visual cortex, they allow computers to "see" patterns through local feature extraction.

1. How Convolution Works

Instead of looking at the whole image at once, CNNs use small filters (kernels) that slide across the pixels. These filters detect edges, then textures, then complex shapes like eyes or wheels as the data moves through deeper layers.

2. Pooling and Subsampling

To reduce computational load and allow the network to be "translation invariant" (recognizing a cat regardless of where it is in the frame), we use pooling layers to downsample the image while keeping the most critical features.

TECHNICAL FACT: Modern CNNs consume up to 90% less memory than fully connected networks of the same depth because of weight sharing across the spatial dimensions.

3. The Rise of Vision Transformers (ViT)

While CNNs are efficient, Transformers are now being applied to images. ViTs treat image patches like words in a sentence, often yielding better performance on massive datasets where global context is more important than local edge detection.