Dual-Initialization Dense Flow and Scatter-Blur-Decay Boundary Correction for Self-Calibrating Aerial Image Mosaicing

This paper presents a complete pipeline for constructing seamless mosaics from pairs of overlapping aerial photographs captured by uncalibrated consumer-grade drone cameras. The system integrates multi-scale normalized crosscorrelation for coarse-to-fine tile registration, six-parameter affine Lucas-Kanade tracking for per-tile local deformation estimation, Brown-Conrady lens distortion fitting via grid-search-initialized weighted least squares, dense Lucas- Kanade optical flow on an 8-pixel grid with iterative dual-initialization refinement, hierarchical Catmull-Rom bicubic upsampling to per-pixel resolution, dynamic-programming seam finding with flow-magnitude-weighted cost, Laplacian pyramid multi-band blending, and a Gaussian-smoothed boundary correction field that eliminates Voronoi discontinuity artifacts. The pipeline introduces several novel methods and combinations: (1) a dual-initialization iterative refinement strategy for Lucas-Kanade flow that competes two independent starting points at each grid node to escape local minima without learned components; (2) error-weighted adaptive spatial smoothing in which Gaussian kernel weights are modulated by inverse photometric warp error rather than image content, so that only high-confidence neighbors influence corrections; (3) hierarchical Catmull-Rom bicubic flow upsampling through three successive doubling stages ( 8 → 4 → 2 → 1), replacing bilinear interpolation to achieve C1-continuous per-pixel flow fields that eliminate period-2 aliasing artifacts; (4) a flow-magnitude penalty term in the dynamic-programming seam cost that steers seams through regions of minimal geometric correction; (5) a scatter-blur-decay boundary correction field that resolves overlap-to-non-overlap color discontinuities without Poisson solves; and (6) self-calibrating Brown-Conrady distortion fitting from a single image pair using tiled NCC residuals as implicit calibration data, with grid-search initialization to avoid local-minimum traps. The complete pipeline is feature-free, requires no calibration targets, GPS metadata, or external libraries, and is implemented as a self-contained application. The following sections describe the mathematical formulation at each stage, document several approaches that were attempted and subsequently abandoned due to specific failure modes, and present the final process flow that produces artifact-free mosaics from DJI drone imagery of natural terrain. The paper concludes with directions for multi-image extension, wide spectral range matching, and integration into signals intelligence (SIGINT) workflows.