Noise may be arisen in the capturing and transmission process of the i mage,which not only corrupts the true information of an image,but also seriously affects the visual effects of the image.Therefore,th e semantic segmentation of noisy images becomes one of the most challenging problems in image analysis.In order to improve the segmentation performance of noisy image,based on the analyses of fully convolutional network (FCN) and the deep convolutional neural network (DCNN),we propose an improved F CN model (IFCN) for semantic segmentation of noisy images.The algorithm introduces a new median pooling method instead of the commonly used max pooling method in convolutional neural network (CNN),which can remove noise and preserve more boundaries information.Our model trains the whole deep network by a direct end-to-end,pixels-to-pi xels mapping way with the back propagation algorithm.By fine-tuning the network structure and adjusting the different parameters,respe ctively,we train and test images with salt and pepper noise and gaussian noise.Experimental results show that the proposed model can achieve a better performance for semantic segmentation of noisy images on Pascal VOC2012dataset,and the accuracy of mean IU is 86.5%, which indicates the effectiveness of the method.