Objective Wavefront sensing of extended targets has been widely applied in object identification, laser atmospheric transmission, laser communication, object imaging, and related fields. In these applications, correlated Hartmann wavefront sensors are used to measure aberrated wavefront phase information in real time. As a key component of adaptive optics systems, the sensing accuracy directly affects the correction capability of the system. In practical sensing, target motion and atmospheric turbulence may cause intensity scintillation, contrast variation, and partial loss of sub-aperture images, resulting in local information loss and reduced accuracy of cross-correlation peak localization. Feature matching provides an effective preprocessing strategy because stable local structures can still be extracted from degraded images to suppress invalid regions before correlation calculation. The speeded-up robust features (SURF) algorithm is robust to scale, rotation, and illumination variation, making it suitable for sub-aperture image preprocessing. However, direct application of SURF in real-time Hartmann sensing is difficult because feature extraction, orientation assignment, and descriptor generation require substantial computation and repeated memory access, especially when multiple sub-aperture image pairs must be processed within one sensing cycle. To address this problem, this study develops a CUDA-accelerated improved SURF algorithm for degraded sub-aperture image preprocessing and introduces a CPU-GPU collaborative framework for multi-image matching.

Methods The proposed method combines CUDA acceleration, descriptor dimensionality reduction, and OpenMP task parallelism. A central sub-aperture image with relatively good quality is selected as the reference image, while the remaining sub-aperture images are treated as target images. SURF matching is performed before cross-correlation calculation so that common valid regions between the reference and target images can be retained and degraded regions excluded from displacement estimation. Several key stages of SURF are parallelized on the GPU. Integral image construction is implemented by parallel prefix summation to reduce redundant operations and improve computational efficiency. Hessian response evaluation at different scales is mapped to two-dimensional GPU thread blocks, enabling parallel box-filter computation for a large number of pixels. Feature point localization is accelerated through parallel comparison of neighboring pixels in both spatial and scale domains, followed by threshold-based refinement to reject unstable extrema. Haar wavelet responses in local circular neighborhoods are computed concurrently, and response vectors are accumulated over angular sectors to determine the principal direction of each feature point. Descriptor generation is further optimized through parallel coordinate transformation and local response computation. To reduce computational cost, the conventional 64-dimensional SURF descriptor is simplified to 16 dimensions. The reduced descriptor retains the central neighborhood around each feature point and divides it into four subregions, preserving representative local structural information while decreasing descriptor construction and matching costs. Because Hartmann wavefront sensing commonly involves multiple image pairs, OpenMP is introduced for CPU task scheduling. Eight host threads are assigned to independent sub-aperture image pairs, while CUDA kernels handle the computationally intensive stages of feature extraction and descriptor generation. This heterogeneous framework improves overall hardware utilization and supports efficient batch matching.

Results and Discussions Experimental results show that CUDA parallelization significantly improves SURF execution efficiency. Compared with the serial implementation, the GPU-accelerated algorithm achieves more than 13 times speedup across all tested image resolutions, while matching accuracy remains above 92%. This indicates that GPU acceleration effectively reduces the major computational bottlenecks of conventional SURF in Hartmann sensing without causing serious degradation in feature correspondence quality. After reducing the descriptor dimension from 64 to 16, the speedup exceeds 15 times relative to the serial implementation, and runtime decreases by approximately 10% compared with the full-descriptor version. Matching accuracy remains above 90% for all tested resolutions. In extended-target Hartmann wavefront sensing experiments, the computational time of the traditional SURF algorithm increases significantly with image resolution, whereas the proposed method achieves an average acceleration ratio of more than 10 times across different sub-aperture resolutions. The sub-aperture slope extraction error experiment further shows that, after feature-matching preprocessing using the proposed method, slope errors at the three tested resolutions are reduced by 54%, 35%, and 22%, respectively, demonstrating improved stability of sub-aperture displacement estimation. Robustness experiments under image degradation and illumination variation further confirm the applicability of the method. As image quality deteriorates, the number of matched feature points gradually decreases, but no false matches are reported in the tested cases and matching accuracy remains 100%. Even at a signal-to-noise ratio of 5 dB, the reduced-descriptor algorithm still detects effective correspondences. Illumination attenuation has a stronger effect on the reduced descriptor; nevertheless, valid matching can still be achieved when the brightness coefficient remains above 0.65, with no mismatched points in successful cases. Matching fails when brightness is reduced to 0.60 of the original image. OpenMP also improves multi-image matching performance. When one reference image is matched with 63 target images, total processing time is significantly reduced compared with host-side serial scheduling, achieving a speedup of more than four times.

Conclusions A CUDA-accelerated improved SURF algorithm is developed for correlated Hartmann wavefront sensing under partial sub-aperture image loss. GPU parallelization is applied to integral image construction, Hessian response computation, feature localization, orientation estimation, and descriptor generation. A 16-dimensional reduced descriptor is used to decrease computational cost, and OpenMP task scheduling is introduced for multi-image matching. The proposed method significantly improves computational speed while maintaining reliable matching accuracy. More than 13 times speedup is achieved after CUDA acceleration, and more than 15 times speedup is obtained after descriptor reduction, while matching accuracy remains above 92% and 90%, respectively. In correlated Hartmann wavefront sensing, the proposed algorithm maintains an average speedup of approximately 10 times across sub-aperture images with different resolutions and stably obtains accurate matching points. The method also shows robustness under image degradation and moderate illumination variation. These results indicate that the proposed method can reduce sub-aperture slope extraction errors while improving the efficiency of feature-matching preprocessing. The CPU-GPU collaborative scheduling strategy also demonstrates practical value for large-scale image matching tasks. This study provides a feasible technical approach for real-time preprocessing of degraded sub-aperture images in adaptive optics systems. Future work can focus on reducing host-device communication overhead and further refining task allocation strategies.