The ratio of the vertical diameter of the optic cup to the optic disc is an important indicator in the clinical diagnosis of glaucoma. To measure the cup-to-disc ratio more accurately, an improved U-shaped convolutional neural network framework was proposed for the problem of segmentation accuracy of the optic disc and optic cup in retinal fundus images. Resnet 34 was used as the new encoder part and a pyramid squeeze attention module at the end of each encoder layer was introduced to extract more valid feature information. A 1×1 convolution was also used instead of 3×3 convolution to simplify the decoding structure, and a 3×3 convolution with a 1×1 convolution structure via a skip connection was used instead of a skip connection. The network model was tested on the DRISHTI-GS dataset after completing training on the in-house dataset, and the segmentation results for the optic disc and optic cup performed 97.61% and 95.32%, 92.91% and 86.75% on the Dice and IOU respectively, demonstrating the good generalization properties of the model.