Using improved density peak clustering algorithm for flower cluster identification and apple central and peripheral flower detection

Abstract

Apple flower detection and positioning are crucial for the mechanical and chemical thinning of flowers, where typically only one or two of the strongest flowers in each cluster are retained. An improved method is proposed that leverages the YOLOv8n model for accurate flower detection. The DPC algorithm is enhanced to automatically determine the number of flower clusters and accurately identify the central flowers within those clusters. To evaluate the performance of the enhanced Single-Layer DPC algorithm, it was compared with several other clustering methods, including DPC, DPC with Shared Nearest Neighbors (DPC-SNN), K-means, K-medoids, Gaussian Mixture Model (GMM), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Spectral Clustering (SC), minibatch and 3W-PEDP. The results demonstrated that the proposed method achieved that the Adjusted Mutual Information (AMI) and Adjusted Rand Index (ARI) were 0.7037 and 0.6043, respectively, on the Flame dataset, surpassing the highest scores obtained by other methods (0.5886 and 0.5116, respectively). Additionally, the improved algorithm reduced the deviation between the clustering center produced by the Single-Layer DPC and the true central flower. Overall, the algorithm effectively reduces clustering center deviations, showcasing its capability to accurately detect and position apple flowers.