Spectral unmixing is an important task for data processing of hyperspectral remote sensing, which is comprised of extracting the pure spectra (endmember) and calculating the abundance value of pure spectra. The most efficient endmember extracting algorithms (EEAs) is designed based on convexity geometry such as pure pixel index (PPI), N-finder algorithm (N-FINDR). Most EEAs choose pure spectra from all pixels of an image so that they have disadvantages like slow processing speed and poor precision. Partial algorithms need reducing the spectral dimension, which results in the difficulty in small target identification. This paper proposed an algorithm that classifies the hyperspetral image into some classes with homogeneous spectra and considers the mean spectra of a class as standard spectra for the class, then extracts pure spectrum from all standard spectra of classes. It reduces computation and the effect of system error, enhancing the speed and precision of endmember extraction. Using the least squares with constraints on spectral extraction and spectral unmixing, by controlling the band average value of the maximum spectral redundant allowance to control the number of endmembers, does not need to reduce the spectral dimension and predetermine the number of endmembers, so compared to N-finder algorithm, such algorithm is more rational.