Computer Vision flashcards, generated for you.

Enable computers to understand and interpret visual information from images or videos, similar to how humans see and make sense of the world.

Computers work with numerical data. Images are represented as grids of pixels, each with intensity values (0-255 for grayscale, or RGB channels for color), which algorithms can then manipulate mathematically.

Sliding a small filter (kernel) over an image to detect local patterns like edges, textures, or shapes by computing weighted sums at each position.

Early layers learn simple features (edges, colors) through many small convolutions; deeper layers combine these into complex features (shapes, parts); final layers recognize whole objects. This hierarchical learning happens via backpropagation updating filter weights.

Pooling (e.g., max pooling) reduces spatial dimensions while preserving important features, making the network more efficient, robust to small shifts, and preventing overfitting.

An activation function introduces non-linearity so networks can learn complex patterns. ReLU (max(0, x)) is popular because it's computationally cheap, avoids vanishing gradients, and empirically works well.

For each position, compute the sum of element-wise products between the kernel weights and the image patch: y = Σ(w_ij * x_ij) + b, where w are weights, x is the input patch, and b is bias.

The chain rule computes gradients of the loss with respect to each weight (∂L/∂w) by propagating errors backward through layers. Weights are updated by gradient descent: w_new = w_old - η(∂L/∂w), where η is the learning rate.

Reusing a pre-trained model (trained on large datasets like ImageNet) and fine-tuning it for a new task. Use it when you have limited data, want faster training, or need good results with fewer computational resources.

Classification: assign one label to the entire image. Detection: locate and classify multiple objects with bounding boxes. Segmentation: assign a class label to every pixel. Increasing complexity and detail in understanding image content.