The excessive reliance on minutiae features in partial fingerprint recognition, leads to degraded recognition performance under conditions of small-area capture and low-quality images. To address the issue, the article propose a partial fingerprint recognition method based on multi-level feature fusion. First, the input fingerprint image undergoes normalization, estimation of the orientation and frequency fields, region segmentation, and Gabor enhancement to improve the stability of subsequent feature extraction. Secondly, during the search phase, a 256-dimensional normalized search vector is constructed using local curvature information, and K-means clustering is employed to filter candidate samples, thereby reducing the computational overhead of large-scale matching. In the fine-matching phase, a local structural model is formulated by combining minutiae information, neighborhood ridge line directions, and local texture information. For each minutiae, four neighborhood descriptor points are sampled, and the number of descriptors per image is controlled between 50 and 200 to balance matching efficiency and feature discrimination capability. Furthermore, dynamic programming, breadth-first search, and linear transformation estimation are integrated to achieve minutiae matching and image alignment. Subsequently, convex and concave ridge shape features are extracted based on ridge width variations, and three-level feature supplementary matching is performed within the overlapping regions. Finally, adaptive weighted fusion of level-2 and level-3 features is performed based on the number of available minutiae, forming a recognition framework characterized by level-1 retrieval, level-2 dominant matching, and level-3 compensation enhancement. Experiments are implemented on the FVC2002, FVC2004, and a self-built fingerprint database. The results show that the proposed method achieves a lower equal error rate (EER) than the comparison methods across different databases and under varying capture conditions. In the smallest capture area scenario, the EER is reduced to 3.35%~6.21%. This method effectively utilizes ridge structure to supplement discriminative information when minutiae are insufficient, exhibiting high recognition accuracy and robustness.