Technical research on image-based virtual scene generation and stitching methods

To address the issues of low matching accuracy and insufficient robustness in multi-source image stitching under complex imaging conditions, this study proposes a new image stitching method. This method integrates multi-threshold local binary patterns with 2DPCA. It first uses the double-threshold MTLBP operator to extract stable local texture features, while simultaneously obtaining the global structural features of the image through 2DPCA. Then, a weighted strategy optimized by grid search is adopted to adaptively integrate the two, constructing an efficient feature descriptor with zero parameters. Experimental results showed that in the ORL face database, when six training samples were used, the algorithm achieved the highest average recognition of 99.2% ± 0.5%. In the Yale face database, under the small sample condition with only 2 training samples, the recognition rate of this method reached 75.0% ± 1.9%. The ablation experiment further quantified the contribution of the fusion strategy. Weighted fusion increased the recognition rate of the Yale library by 1.4% compared with the simple concatenation of features, and improved the PSNR index of the MSU Panorama dataset by 0.6 dB, confirming the effectiveness of the proposed strategy. The results show that the proposed fusion strategy not only enhances the discriminative power and robustness of image description through feature complementarity but also achieves a balance between accuracy and efficiency.