With the aim to predict the wheat yield accurately, an improved wheat spikes detection method based on feature pyramid network was proposed. In order to solve the problem of misdiagnosis or omission in the detection results, channel attention mechanism and spatial attention mechanism were introduced into the coding and decoding regions of the original feature extraction network, which increased the extraction of spatial information and semantic information on the wheat spikes and effectively improved the detection performance of the network for obscured wheat spikes. At the same time, a weighted-region proposal network was designed to improve the input of the original region proposal network, in which several low-resolution feature maps with strong semantic information characteristics were fused together on channel levels. After a series of full connection layers and activation functions, the fused feature map was converted to probability of the corresponding channels, which were used to weight the underlying high-resolution feature maps to enhance useful information channels. Thus, a more accurately detection frame was generated for smaller spikes which were difficult to detect. The experimental results of the collected wheat spikes images showed that the method could significantly improve the detection effect of the shaded and smaller wheat spikes, where the precision of recognition, recall rate and average precision were 80.53%, 87.12% and 88.53%, respectively. Through the comparative analysis of wheat spikes detection results in different periods on the public ACID data set, the validity of the proposed method was further verified.