In order to solve the geometric distortion caused by filter change in the imaging process of multi-channel spectral data, a multi-channel spectral image registration method based on speed up robust features (SURF) and maximum submatrix is investigated. The features of the multi-channel spectral images are extracted by the SURF algorithm, and the preliminary registration images are obtained through perspective transformation. In addition, aiming at the problem that there exists the invalid regions with zero pixel value on the image edge after registration, a method for detecting the largest inner rectangle in images by maximum submatrix is proposed to remove the invalid edge regions and maximize the retention of valid area information. The multi-channel imaging data of the murals are tested. The results show that the proposed method has better robustness to image scale and brightness variation, and can avoid the influence of invalid regions on the subsequent spectral reconstruction and color restoration occurred in the other registration methods. It also has better performance in registration accuracy, information maximization retention, and time efficiency.