Factors such as illumination or imaging conditions can cause a nonlinear grayscale difference between images, resulting in poor matching of images. To solve this problem, this paper proposes a new image registration algorithm based on the stacked autoencoder (SAE) network and local binary pattern with a circular and linear neighborhood (CL-LBP). First, the CL-LBP feature descriptor is extracted and matched by combining the improved local texture operator with the regional feature. Then, the supervised learning classification method is used to eliminate mismatches. Finally, the constructed matching representation is trained using the SAE network and the depth features of the matching representation are extracted and connected to a logistic classification layer to classify the matching pairs. The experimental results show that the algorithm has good matching accuracy in matching the nonlinear grayscale difference images. Moreover, it has a good matching effect in the actual sea ice images.