Abstract: Misalignment degrades the performance of inspection algorithms. Misalignment problem of solder joint inspection is investigated quantitatively through evaluating the current algorithm's sensitivity to misalignment. A new learning method is proposed in which a set of virtual samples generated by perturbing the original samples on position, angle and scale are used for training. Synthetical false rate is defined to evaluate the performance of the inspection algorithm regarding misalignment. The experiment results showed that the proposed method is robust to the misalignment while the time cost remains unchanged.