A multi-class image segmentation method based on multi-task learning is proposed for diabetic retinopathy (DR) images. Specifically, the dominant pixels without lesion information are removed by the Otsu thresholding algorithm; subsequently, the image is segmented into several small-sized images by the method of sliding window segmentation to solve the problems that the resolution of medical images is too large and the proportion of lesions in the image is small; then, sub-images without lesions are eliminated to increase the proportion of those with lesions; finally, multi-output multi-lesion image segmentation is performed by leveraging the multi-task learning properties of UNet++ and replacing traditional upsampling with transposed convolution. When the proposed method is verified on the international public Indian diabetic retinopathy image dataset (IDRID) and dataset for diabetic retinopathy (DDR), it achieves a mean area under precision-recall curve (mAUPR) of 0.7131 on IDRID and an mAUPR of 0.5691 on DDR.