Abstract: Maize leaf rust is a destructive foliar disease that reduces the effective photosynthetic area and threatens stable maize production. Accurate lesion segmentation is essential for objective severity assessment, yet field images contain small lesions, blurred boundaries, uneven illumination, complex backgrounds, and dense leaf textures. This study developed a lightweight edge-guided segmentation network (LESNet) to improve lesion delineation while satisfying the computational constraints of mobile devices. LESNet adopted DeepLabV3+ as its encoder-decoder framework and used a modified MobileViTv2 backbone for efficient local and global feature representation. Parallel depthwise separable and atrous depthwise convolutions enlarged the effective receptive field while preserving local details. The hidden dimension of the feed-forward network was reduced from four times to twice the input channel dimension to decrease computational redundancy. An adaptive graph channel attention (AGCA) module integrated average- and max-pooled descriptors and modeled local and learnable inter-channel relationships. An edge-guided attention (EGA) module generated multi-scale Gaussian edge cues and fused them with semantic features from Stages 2, 3, and 4. The atrous spatial pyramid pooling (ASPP) module was further improved through cross-branch feature fusion and adaptive reweighting to strengthen multi-scale semantic interaction. The self-built dataset contained 743 pixel-level annotated field images, comprising 372 training, 297 validation, and 74 test images. Data augmentation increased the training set to 1,114 images. Annotations distinguished background, healthy leaf tissue, and rust lesions. Evaluation combined class-wise IoU, mIoU, mPA, parameter count, floating-point operations, inference time, and frame rate to examine both segmentation quality and deployment efficiency. All comparison models were evaluated under identical experimental settings, and each experiment was independently repeated three times. Mobile inference was assessed using Open Neural Network Exchange (ONNX) models integrated into the same application on a Vivo Y300 Pro and a Vivo S15. On the self-built dataset, LESNet achieved a mean intersection over union (mIoU) of 82.41%, a mean pixel accuracy (mPA) of 89.65%, and a lesion-class IoU of 65.16%. The background- and healthy-leaf-class IoU values reached 91.64% and 90.42%, respectively. Its mIoU exceeded those of UNet, PDSNet, SegFormer, and BiSeNetv2 by 1.99, 5.78, 6.19, and 3.37 percentage points, respectively. Stratified evaluation yielded mIoU values of 71.05%, 79.80%, and 82.64% for mildly, moderately, and severely infected samples. The corresponding lesion-class IoU values were 29.94%, 54.35%, and 69.71%. LESNet ranked first in mIoU at all three severity levels, exceeding the second-best model by 1.47, 0.80, and 1.94 percentage points. The lower accuracy for mildly infected samples indicated that small, sparse, and low-contrast lesions remained the primary source of error. In contrast, LESNet retained dense lesions and irregular boundaries more effectively as disease severity increased. The model contained 0.86 million parameters and required 6.33 GFLOPs for a 512 × 512-pixel input; both measures were the lowest among the compared models. Its mean inference time on the two mobile devices was 99.23 ms per image, equivalent to 10.05 frames per second. The comparison models required 139.81 to 546.56 ms per image and achieved frame rates of only 1.85 to 7.14 frames per second. Ablation experiments confirmed complementary contributions from the modified MobileViT block, AGCA, improved ASPP, and multi-stage EGA. Applying EGA at all three stages increased mIoU from 81.71% to 82.41% while adding only 0.04 million parameters. An additional evaluation on 588 images from a public multi-disease leaf segmentation dataset produced an mIoU of 70.60% and an mPA of 78.61%. Both values were higher than those obtained by PDSNet, SegFormer, BiSeNetv2, and LinkNet under the same evaluation protocol. Lesion-response visualizations further showed fewer false responses and more concentrated activation in lesion regions under strong illumination and shadow. These results demonstrate that LESNet balances lesion segmentation accuracy, model compactness, and mobile inference efficiency under the evaluated field conditions. The remaining limitations include the single-region dataset, limited independent repetitions, absence of significance testing, and lower lesion accuracy than background and healthy-leaf accuracy. Future work should incorporate multi-region and multi-disease data, repeated random seeds, statistical testing, model compression, and device-specific inference optimization.