Computer Vision flashcards, generated for you.

Enable computers to understand and interpret visual information from images and videos, mimicking human visual perception to extract meaningful data about what is shown.

A 2D array (matrix) of pixel values, where each pixel has coordinates (x, y) and intensity/color information. Grayscale images have 1 channel; color images have 3 (RGB) or 4 (RGBA) channels.

Filters detect local patterns (edges, textures, shapes) by performing convolution. A kernel slides over the image, multiplying overlapping regions element-wise and summing the result to highlight specific features.

It measures similarity between a kernel (learned feature pattern) and each patch of the image. High values indicate the patch closely matches the kernel's pattern; this creates feature maps showing where patterns occur.

(I * K)[i,j] = Σ Σ I[i+m, j+n] · K[m,n], summed over all kernel positions. In CNNs, this is applied with learnable weights that the network optimizes via backpropagation.

Early layers learn simple features (edges, corners); middle layers combine these into textures and shapes; deep layers recognize objects and semantics. Each layer's output becomes the next layer's input, creating a hierarchical feature pyramid.

Pooling (max, average) downsamples feature maps to reduce computation and parameters, while retaining dominant features. It provides translation invariance—slight shifts in input produce similar outputs—making the network robust.

Classification: assign one label to entire image. Detection: localize and label objects with bounding boxes. Segmentation: label every pixel with a class (semantic) or instance ID (instance).

A 2D activation grid where each cell's value indicates how strongly that spatial location matches a specific learned feature pattern. Multiple feature maps per layer capture different patterns (edges at different orientations, textures, shapes).

Attention allows each pixel to directly compare and learn relationships with distant pixels globally, whereas convolution only sees local neighborhoods. This captures long-range dependencies more efficiently and achieves state-of-the-art results on many tasks.