Because the result of method used by the project group currently which is a combination of floodfill and template matching is poor, and there is also under-segmentation or over-segmentation, this paper proposes the application of fully convolutional networks in semantic segmentation for wheat images. Firstly, the output information of the second pool layer is integrated as the input of the Softmax layer. Then, the Batch Normalization layer is introduced into the network layer, and 21 classes of output of the network are changed into the output of the 2 classes because of the characteristics of wheat. And the paper uses the F-measure to evaluate the result. The experimental results show that the proposed network can improve the segmentation result.