Abstract: Virtual test line for two feature extraction threshold is set manually adjust the defect, in order to ensure long-term system in unattended mode high-precision and stable operation, proposed a BP neural network based adaptive threshold technology to solve the threshold selected issues.The use of forward propagation BP network operating signal, error signal back-propagation characteristics, levels of connection weights automatically correct, the resulting error is very small output.Experimental results show that the algorithm can be obtained in the actual application environment more appropriate threshold.