The relationship between underground salt bodies and oil and gas reservoirs is inseparable. The accurate recognition of salt bodies is of great significance to the oil and gas reservoir exploration and drilling path planning. In existing deep learning methods，the size of the receptive field is unchanged，and the convolution kernel cannot be dynamically adjusted for the feature capture according to the size of the salt body in a seismic image. As a result，the part of the global information is ignored，which results in poor recognition at the boundaries or narrow areas of the salt bodies. In response to the above problems，this paper proposes a new method based on U-Net with SKNet as the encoder to extract salt body features.It has a dynamic selection mechanism that allows the sizes of the receptive fields to be adjusted adaptively according to multiple scales of the input information. In addition，it combines the position and channel self-attention mechanism and the hyper column method for feature fusion. The improved U-Net method is used to evaluate the TGS salt body data set. The recognition results have an intersection over union（IoU）of 85.66% and a pixel accuracy of 96.1%.