In the process of image acquisition, the image often contains certain noise information, which will destroy the texture structure of the image and thus interfere with semantic segmentation tasks. Most of the existing semantic segmentation methods based on noisy images adopt models featuring first denoising and then segmentation. However, they often lead to the loss of semantic information in denoising tasks, which thus affects segmentation tasks. To solve this problem, this study proposes a multi-scale and multi-stage feature fusion method for semantic segmentation of noisy images, which uses the high-level semantic information and low-level image information of each stage in the backbone network to enhance the semantic information of target contours. By constructing a staged collaborative segmentation denoising block, collaborative segmentation and denoising tasks are iterated, and then more accurate semantic features are captured. In addition, quantitative evaluation is carried out on PASCAL VOC 2012 and Cityscapes datasets. The experimental results show that the model still achieves positive segmentation results under the noise interference of different variances.