Convolutional neural network has made great breakthroughs in the field of computer vision. With its powerful image processing ability,we transform classification of underground sedimentary salt-body into image semantics segmentation problem. Deep convolution neural network is applied to implement pixel-level semantics segmentation of salt seismic images. We increase depth of the network and adds batch normalization and Dropout processing based on U-Net, which makes the neural network model with higher reliability and stronger generalization ability. With experiments, it was found that adding batch normalization layer after convolution layer, and adding Dropout after the pooling layer and concatenate layer improve steadily segmentation performance of the model for salt-body seismic image.