Abstract: Objectives: Large-swath side-scan sonar (SSS) image stitching is essential for seafloor geomorphological surveying and marine engineering applications, yet is severely challenged by underwater acoustic noise, speckle interference, and geometric distortion from towfish motion, yielding sparse and unstable feature points. Conventional scale-invariant feature transform (SIFT), speeded up robust features (SURF), and Accelerated-KAZE (AKAZE) pipelines require exhaustive manual parameter tuning including feature thresholds, random sample consensus (RANSAC) inlier thresholds, and so on, which is labor-intensive, error-prone, and strongly operator-dependent. To eliminate manual intervention while preserving high registration accuracy, we propose an intelligent framework integrating spatial gradient fuzzy C-means (SPGFCM)-AKAZE regional feature extraction with Tree-structured Parzen Estimator (TPE) Bayesian parameter optimization for fully automated large-swath SSS imagery stitching. The objective is to achieve autonomous, high-precision, and computationally efficient stitching of massive strip sonar datasets without any human-in-the-loop parameter adjustment. Methods: The framework comprises four synergistic stages. First, preprocessing and partitioning: raw sonar strips undergo radiometric correction, geometric rectification to establish a unified coordinate system, then are partitioned into manageable blocks for parallel processing. Second, SPGFCM-AKAZE regional feature extraction: each block is clustered by SPGFCM into K texturehomogeneous regions based on spatial gradient magnitude; AKAZE features are independently extracted within each region via fast explicit diffusion (FED) in nonlinear scale spaces, guaranteeing robust keypoints even in weak-texture sediments. Third, multi-level constrained matching and Bayesian auto-tuning: Lowe's ratio test, DBSCAN clustering, and RANSAC geometric verification progressively filter mismatches. A TPE Bayesian optimizer automatically tunes critical parameters including AKAZE Hessian threshold, Lowe's ratio, DBSCAN radius, and RANSAC threshold, guided by a composite quality score function fusing node count, inlier ratio, spatial distribution entropy, and matching uniqueness. A global prior warm-start mechanism reuses converged parameter distributions across adjacent blocks, reducing search space and accelerating convergence. Fourth, global transformation and seam-line blending: iterative global transformation with sigma clipping and RANSAC refinement yields a unified projective model; dynamic programming seam-line detection and distance-weighted linear blending produce seamless stitching. Results: Validation was performed on real Klein 3000 dual-frequency SSS data (100/500 kHz, 400 m swath, 0.1 m/pixel) collected in Dapeng Bay, Shenzhen, comprising four survey lines with 142, 156, 114, and 140 images and approximately 50% overlap. The proposed method achieves a standard point error of 1.30 m, PSNR of 20.04, and SSIM of 0.8421, comparable to or better than manually tuned SIFT, SURF, and AKAZE baselines. Ablation studies demonstrate that removing SPGFCM increases standard point error by 0.30 m (from 1.37 m to 1.67 m), confirming that regional texture-guided extraction is indispensable. Removing the Bayesian optimizer and reverting to fixed parameters degrades both accuracy and efficiency. The TPE framework reduces total matching time by 68.3% versus grid search and 72.0% versus random search; per-block inference converges to 0.8-1.0 s after warm-start, far below the 20-28 s of exhaustive search. TPE exhibits rapid convergence within 7-10 trials with stable composite scores around 0.110, demonstrating efficient knowledge reuse without overfitting. Conclusions: The proposed framework successfully automates the entire SSS image stitching workflow without any manual parameter adjustment. By synergizing SPGFCM-guided regional feature enhancement with intelligent Bayesian parameter search, the method delivers high-precision registration and markedly improved computational efficiency for massive strip datasets, satisfying practical engineering requirements for automated marine surveying. Future plans include further testing with images from more diverse survey areas and devices to validate the generalization capability of the method, as well as improving its efficiency.