To improve the quality of lesion segmentation in leaf images with clutter background and uneven lighting, a novel method was presented based deep salient object detection. Saliency target detection network was used to generate the saliency map of grape disease leaf image. The local and global information of the image were extracted by multi-resolution grid structure and fused into prediction features. Then, the diseased areas on the leaves were segmented by the adaptive threshold method on the significance map of diseased leaves, and the post-processing was carried out by the morphological method. The segmentation experimental results show that, on the test set A, the Matthews correlation coefficient(MCC) of the proposed method is 0.625, which is slightly lower than 0.689 for the convolutional neural net-work comparison algorithm(FCN) of. On the test set B, the MCC for the proposed method is 0.338, much higher than 0.072 for the FCN. It shows that the proposed method has good balance between the segmentation accuracy and gener-alization.