NeuQuant

NeuQuant is a neural-network based color quantization algorithm developed by Anthony Dekker for reducing the colors in digital images. It is widely used in GIF encoding, where a palette of up to 256 colors is required. The method is based on a Kohonen self-organizing map that operates in RGB color space. The algorithm maintains a fixed-size map of 256 neurons, each holding a color vector.

During training, each input pixel color is compared to all neuron weight vectors to find the best matching unit (BMU). The BMU and a neighborhood of adjacent neurons are adjusted toward the input color by a learning rate that decreases over time; the size of the neighborhood also decays. This cooperative adjustment causes the neurons to organize into a palette that represents the prominent colors in the image. After training, the palette is formed from the neuron weights, and each input pixel is remapped to its nearest palette color. Dithering can be applied to reduce quantization artifacts.

NeuQuant offers a balance between color fidelity and computation speed, and is capable of producing perceptually pleasing palettes for natural images. It can be configured with speed-oriented options, such as reduced sampling of pixels, and different initial radius and learning-rate schedules. It is implemented in various image-processing libraries and tools, and remains a reference approach to neural-network based color quantization.