Due to continuous pooling and down sampling, the resolution of the final feature hotspot map is seriously lost in the traditional full-convolution neural network, which leads to the loss of detail characterization ability of segmentation results. In order to make up for this defect, the feature map of middle layer is often fused by jumping connection to restore spatial information. Due to the failure to make full use of the low-level feature information of the network, the feature fusion stage of the traditional full-convolution network has some defects. This study makes an in-depth analysis of this phenomenon. The feature information enhancement method based on feature pyramid is adopted before the upper sampling path to overcome the deficiency of semantic information of shallow feature graph, so that the entire network can make full use of the feature graph generated by forward calculation and improve the segmentation result. The algorithm proposed in this study achieves 75.8% average pixel accuracy and 83.9% weight frequency crossover ratio on the Pascal VOC data set, effectively improved the classification accuracy.