In remote sensing devices, image rotation has become a crucial factor affecting imaging results, but if not corrected, it will lead to off-axis distortion of detected target information, preventing the acquisition of accurate azimuth information. Current image rotation correction methods mainly include optical de-rotation, mechanical de-rotation, and digital de-rotation. However, both optical and mechanical de-rotation methods require the addition of new devices to the original imaging system, imposing high demands on device weight and motion accuracy. Therefore, we propose a digital de-rotation algorithm. Meanwhile, the large field-of-view infrared images are advantageous for obtaining abundant terrain information. Thus, it is necessary to stitch the corrected rotated images after correction. However, there is currently no well-established solution for the challenging task of stitching rotated and corrected images. Existing image stitching methods demand high image quality and a significant number of matching points between images. The overlapped areas between the rotated images acquired by the detector are typically small to expand the field of view. Thus, it is essential to develop a stitching algorithm specifically designed for rotated and corrected images.

We propose a two-dimensional pointing mirror rotation correction and stitching method. Firstly, the image rotation correction algorithm is based on the optical imaging principles of the detector. It builds the imaging model of the two-dimensional pointing mirror, as shown in Eq. (5). Subsequently, the image rotation correction method is derived by reverse deduction of this model, as shown in Eq. (10). Then, the stitching algorithm for the image rotation-corrected images is shown. This method relies on a simulated field-of-view model based on information such as the elevation angle, azimuth angle, and detector specifications of the two-dimensional pointing mirror (Fig. 4). By employing the model to determine the pixel relationships between models, the positional information between images is obtained. Subsequently, based on the orientation information and pixel relationships among images, the image stitching results are achieved.

To validate the effectiveness of the proposed image rotation correction and stitching algorithm, we collect a set of real image rotation data using the detector from our research group. The experimental results indicate that our image rotation correction method can eliminate image rotation errors, and it exhibits an 8% improvement in time efficiency compared to the correction methods in previous studies. The stitching results demonstrate that the proposed image rotation correction algorithm is not constrained by the size of the overlapping area between images or the image quality. Additionally, it achieves seamless and natural large-field-of-view stitching results. In comparison to more advanced stitching algorithms currently, this method is simple and fast and produces tightly-knit and natural stitching results. The contrasted algorithms under small overlapping areas fail to yield correct stitching results. Meanwhile, if the pitch and azimuth angles of the detector are fixed, the calculated pixel relationships between the stitched images can be directly applied to the stitching task, enabling real-time stitching in space.

We propose a method for image rotation correction and stitching in response to the image rotation distortion caused by two-dimensional pointing mirrors and the blank space in the field of image rotation image stitching. Additionally, field experiments are conducted using our research group’s detector to validate the effectiveness of the proposed image rotation correction algorithm and image stitching algorithm. Meanwhile, a set of nine-grid image rotation data is collected. Experimental results demonstrate that the proposed image rotation correction algorithm successfully corrects distorted images caused by image rotation and improves correction efficiency. It is not influenced by the overlap area size between images and image quality, and can accurately complete the image stitching task, leading to naturally seamless images with almost imperceptible seams. The proposed algorithm performs well under small pointing mirror installation error, and the detector’s optical distortion is minimal. However, for situations with significant installation errors or substantial optical distortion in the detector, the modeling process should consider the installation error matrix and optical distortion. Therefore, adjustments to the proposed correction method should be made based on the characteristics of the employed specific detector in the further research.