Traditional U-Net semantic segmentation model has a wide range of applications in the field of biomedical image, but the accuracy of the model is limited by the single-scale and the loss of information caused by up-sampling and down-sampling. To solve the above-mentioned problem, this paper proposes a high level and low level information-rich multi-scale biomedical image segmentation algorithm based on U-Net Encoder-Decoder structure and atrous separable convolution. The proposed algorithm consists of feature extraction network and multi-scale semantic segmentation prediction network. The feature extraction network uses atrous separable convolution and similar residual blocks to replace the up-sampling, down-sampling and convolution blocks in the original U-Net respectively, which maximizes the retention of information while increasing the perception field of vision. Channel attention mechanism is proposed to highlight the expression of core features and suppress the expression of unrelated region of background. The convolutional features with image-level features encoding global context are data mined at multiple scales to boost the performance further. Simulation experiments are conducted on collected embryos and DRIVE datasets, the experience results show that the proposed method has higher accuracy and robustness compared with U-Net and its derivative model.