Coarse-to-Fine Pyramid Feature Mining for Wheat Head Detection

Automatic detection of wheat head attracts extensive attention for efficient and effective wheat farm management and breading study. Benefiting from the powerful learning capability of deep convolution neural networks (DCNNs), recent work have demonstrated the potential and feasibility of the detection automation for wheat head. However, because of the appearance uncertainty of wheat head and the large variability of imaging conditions, detection performance is still needed to be improved for real application. This study exploits a novel coarse-to-fine pyramid feature mining network (CFPFM-Net) for anchor-free wheat head detection. The proposed CFPFM-Net incorporates pyramid context fusion and a U-net-based refining module for coarse-to-fine feature mining, and only simply predict the centerness and the size of wheat head based on the aggregated context in a high resolution permitting accurate detection for small-size wheat head. Moreover, to capture more effective features in various scales and alleviate the gradient vanishing problem, we leverage the auxiliary supervision on several intermediate feature maps in the training phase. Experiments on the Global Wheat Head Detection (GWHD) dataset have demonstrated that the proposed framework achieves superior performance over the existing state-of-the-art methods.