An Anchor-Free Refining Feature Pyramid Network for Dense and Multioriented Wheat Spikes Detection Under UAV

The number of wheat spikes is a crucial index for evaluating the yield, and the precise detection of wheat spikes in an image plays an important role. Among various methods, deep learning-based approaches show impressive results in the task of wheat spike detection. However, the precise detection and recognition of wheat spike encounters large challenges due to complicated backgrounds, arbitrary orientations, and dense distribution in wheat spike images. To alleviate these issues, we have developed an anchor-free refining feature pyramid network (AFRFPN) that gets rid of horizontal bounding boxes (HBBs) from the network. First, the refining feature pyramid network (RFPN) has been introduced into extract richer features of wheat spike with highly variant appearances and multiple scales. Then, learning from the idea of coarse-to-fine, the two-stage anchor-free oriented detection (AFOD) module has been designed. The AFOD module first generates a set of coarse detection (CoDet) results in the way of anchor-free and then further fines them to achieve high-quality predicting bounding boxes (BBs). The number of wheat spike images is insufficient, resulting in poor performance of wheat spike detection modules. To mitigate the lack of the data in the task of oriented wheat spike detection, based on the global wheat head detection (GWHD) dataset, we released a new large-scale wheat spike dataset by relabeling the samples, termed it as rotated GWHD (R-GWHD) dataset. Massive experiments show that the proposed method can achieve 90.6% mAP and 96.7% recall, outperforming other state-of-the-art methods. Additionally, the experiments related to the counting of wheat spikes have been conducted, showing that the developed module can achieve the MAE of 4.95 and RMSE of 7.68, which demonstrates the excellent performance of the proposed method.