Improved P2PNet with Multi-Scale Feature Fusion and Point-Relation Refinement for Image-Based Maize Kernel Counting

Accurate counting of visible maize kernels from RGB images is important for maize phenotyping and yield-related trait assessment, but remains challenging because kernels are densely distributed and frequently affected by occlusion and local appearance similarity. Point-based counting provides a promising alternative to bounding-box-based detection by reducing annotation requirements while retaining object-level localization. However, the original P2PNet relies on a single prediction feature with limited direct interaction between high-resolution spatial information and high-level semantic features. To address this limitation, a bidirectional feature pyramid network (BiFPN) was incorporated into P2PNet to enable bidirectional multi-scale feature fusion among $C_3$, $C_4$, and $C_5$ while retaining $P_4$ as the prediction level. A Point Relation Refinement Module (PRM) was further investigated to examine the effects of candidate-point relational modeling on initial predictions through multi-head self-attention and residual refinement. Experiments were conducted on a maize kernel dataset containing 452 RGB images and 92,314 point annotations. Across three random seeds, the combined BiFPN+PRM model achieved the lowest mean absolute error (MAE) of $26.33\pm 1.78$ and mean relative error (MRE) of $10.50\pm 0.48\%$ among the four main P2PNet ablation configurations, compared with $27.70\pm 1.92$ and $11.54\pm 0.83\%$ for the original P2PNet. BiFPN alone did not consistently improve counting accuracy but reduced computational complexity from 117.474 to 105.159 GFLOPs and increased inference throughput from 279.98 to 288.90 FPS. PRM exhibited sample-dependent behavior and substantial computational overhead. However, top-512 candidate selection reduced inference latency by 74.44% compared with global attention under the actual-resolution evaluation protocol, with slightly lower mean MAE and RMSE but a higher mean MRE (10.50% to 11.01%). These findings demonstrate the potential of the proposed framework for non-destructive image-based kernel quantification in maize phenotyping.