The reference based Non-Uniformity Correction (NUC) is susceptible to the special application situation and the working condition drift of the instrument. This may result in the output images with poor quality. To solve this problem, a scenebased NUC algorithm for the total variation model is realized on the Field Programmable Gate Array (FPGA) platform. The application principle of the model and some measures for suppressing image ghosts and degradation are presented. Function debugging and performance testing are completed in some cooled LWIR detectors. The experimental results show that this algorithm can accomplish correction convergence in 20 frames and satisfactory imaging quality is achieved.