Aiming at the high false detection rate due to the image translation and rotation, the low efficiency and low reliability of manual inspection, and the impact of image registration precision on the detection accuracy, a method based on Fourier-Mellin transform for liquid crystal display (LCD) defect detection is proposed. The basic fundamental of the method is as follows. The rough matching based on the Fourier-Mellin transform is performed between the detected image and the standard image firstly, and then the fine matching based on the speed-up robust features/scale-invariant feature transform (SURF/SIFT) algorithm is performed. The weighted average fusion is used for the standard image and the image after registration to obtain the final image. The LCD defects are inspected by local adaptive threshold segmentation and subtraction method, with the defect location and information marked. The experimental results show that the proposed method is robust to image translation and rotation, and can effectively detect the LCD defects with an accuracy rate up to 98.667%.