Phase congruency image mosaicking approach for aerial mid-wave infrared low-overlap array scanning images

Abstract

With the rapid development of imager manufacturing technology, mid-wave infrared (MWIR) array scanning images have been widely used to embody abundant thermal radiation geographic information. Due to the limited field of view (FOV) of the MWIR imaging detector, image mosaicking is essential for combining multiple overlapping images into a larger FOV image. However, MWIR images simultaneously suffer from poor image quality and a low signal-to-noise ratio (SNR), presenting significant challenges to existing mosaicking methods, particularly under low-overlap conditions. To overcome these challenges, this study proposes a robust phase congruency (PC) image mosaicking approach for aerial MWIR array scanning images based on image positions derived from Position Orientation System (POS). First, a joint corner-edge PC (CEPC) feature detection strategy is implemented to enhance feature point detection in MWIR images. Subsequently, a fractional average PC localization and orientation histogram (FAPC-LOH) descriptor is developed to generate robust feature descriptors. Additionally, image pairs and matching correspondences within overlapping regions are filtered using the initial image positions to prevent mismatches and ensure the reliability of feature points. Valid feature points are incorporated into the global consistency rectangling alignment model based on topology analysis to obtain the rectangular mosaicking results. Finally, ground control points (GCPs) are used to correct the planar projection error of the mosaicked images. The proposed mosaicking method is rigorously evaluated on six MWIR datasets collected from three cities, encompassing diverse scenarios, flight altitudes, imaging times, and overlap rates. Results demonstrate that our PC-based approach improves the mean number of inliers (MNI) by 5–12 times, increases the rate of successful matching (RSM) by 21.2–46.93% with an average RSM of 99.74%. It also achieves an average alignment root mean square error (RMSE) of 2.11 pixels and an average geometric positioning accuracy of 1.26 m (RMSE) across six datasets. Furthermore, the alignment results outperform those of representative mosaicking algorithms and popular commercial software, achieving superior global and local alignment along with enhanced positioning accuracy.