CRFormer: Cross-Resolution Transformer for segmentation of grape leaf diseases with context mining

In the smart agriculture community, automatic segmentation is an important basis for plant disease detection and identification. However, the complex background and texturally rich edge detail make it difficult to segment grape leaf disease. The existing methods seldom consider the in-depth understanding of the whole scene that is helpful for the precise segmentation of small diseased regions. To this end, we build three datasets and propose a tailored segmentation architecture referred to as the Cross-Resolution Transformer (CRFormer) for field grape leaf disease. Concretely, we introduce a large-kernel mining (LKM) attention operation to reshape the weight matrix, which can adaptively encode channel and spatial information for small disease areas with complex backgrounds. Furthermore, we design a multi-path feed-forward network (MPFFN) to further mine different scales of contextual information by applying convolutional pairs. Besides, CRFormer leverages a lightweight decoder to improve the ability of multi-scale information aggregation. Extensive experiments have demonstrated that CRFormer remarkably outperforms leading methods on the datasets we built, including Field-PV, Syn-PV, and Plant Village. Our CRFormer achieves 88.78% IoU with less computation than competitors on the Field-PV dataset. The ablation experiments investigated the effectiveness and robustness of the core proposed components in CRFormer.