The original U-Net integrates a jumping structure with high-level and low-level image information, which makes the U-Net model perform well in segmentation, but the results still present poor segmentation, over-segmentation, and under-segmentation at the edges of cervical nucleus. Then an improved U-Net network for image segmentation is proposed. First, the densely connected DenseNet is introduced into the encoder of U-Net to solve the problem that the encoder is too simple to extract abstract high-level semantic features. Then different weights are given to the cervical nucleus nuclei and background in the binary cross-entropy loss function, so that the network pays more attention to the learning of nuclear characteristics. Finally, during the pooling operation, reasonable weights are assigned to the pixel values in the pooling domain to avoid losing information in the pooling layer. Experimental results reveal that the improved U-Net network can behave better in cervical cell segmentation with a more robust model, and the proportions of over-segmentation and under-segmentation are also smaller.