The present application relates to an image segmentation method and device operating on an electronic device. The image segmentation method comprises the following steps: performing deep learning to obtain an FCN (Fully Convolutional Network) model, and calculating the loss using L(pji)=−(1−pji)r log(pji) in the deep learning process; inputting an image to be segmented to the last updated FCN model to obtain the probability that each pixel in the image to be segmented is in each of the categories, and selecting the category corresponding to the maximum probability as the category determined by the image segmentation for the pixel. By improving the loss function of the FCN model, the accuracy of image classification is improved, and chart information in an electronic document is accurately extracted by means of pixel classification.