Short circuit is a key factor which drastically affects the efficiency of metal electrorefining. Infrared image of the intercell busbar region is used to perform short circuit detection. To cope with the high thermal background, a two-level short circuit detection method is designed. Firstly, with background subtraction, high intensity short circuit electrodes, as well as the background, are removed, and normal working electrodes are preserved. In the second stage, suspicious short circuit areas are sifted out by normal electrode detecting and texture period estimation. Gaussian difference filter (DoG) which is based on the human visual system is improved to match the target gray distribution. A comparative experiment indicates that the proposed orthogonal DoG outperforms the original DoG and top-hat in the accuracy of normal electrode detection. The two-level detection method in this paper is applied in a copper electrolysis plant and exhibits superiority in locating short circuits and avoiding miss detection.