An embedded system for grading of pre-harvest cotton was designed on overall. ARM (TMS320DM365) was regarded as main processor of the embedded terminal, and DSP (TMS320DM6437) was regarded as slave processor. Linux system environment was built on ARM, and video capture and man-machine interaction were executed in parallel. Simulation video flows were captured by camera and decoded in decoder chip (TVP5146). Digital video flows were displayed real-timely on user interaction interface and the video process request was sent to DSP by SPI. After DSP received the request signal during video flows capture synchronously, the current frame was segmented and discriminated by features extraction, and the result was returned to ARM. Field experiment showed that the system could recognize the grades of raw cottons quickly with an accuracy of 90.48%.