Abstract
Accurate counting of spikes in crested wheatgrass, an important forage resource, is essential for breeding and yield evaluation. However, traditional manual counting is inefficient, and instance-level supervised methods face challenges such as high annotation costs and counting errors caused by overlapping targets in complex field scenes. To address these issues, this study proposes the Multi-Granularity Gating Image-Level Supervision Count Network (MGG-ISCNet), a lightweight image-level supervised counting network. The network integrates multi-granularity features adaptively and employs a lightweight regression head with two 1D convolution layers and global average pooling for efficient feature compression, greatly reducing parameter complexity. Requiring only image-level count labels without positional annotations, the proposed approach substantially lowers labeling costs. On a self-constructed crested wheatgrass dataset, the MGG-ISCNet achieved an MAE of 2.73, RMSE of 3.86, and R2 of 0.81. Furthermore, transfer experiments on the wheat spike dataset GWHD2020 demonstrated strong generalization. The proposed method achieved the best accuracy among both instance-level and image-level supervised approaches, with MAE = 3.63, RMSE = 4.73, and R2 = 0.95, while featuring significantly fewer parameters (61.08 M) compared to the existing image-level method. Overall, this work provides an efficient and lightweight solution for spike counting in crested wheatgrass and other cereal crops, offering valuable support for breeding and forage production.