Lightweight multi-scale aggregation network with full depthwise separable convolutions for Panax notoginseng disease segmentation

As a crucial medicinal and economic crop, the identification of diseases in Panax notoginseng is essential for ensuring yield and quality. However, existing deep learning models face challenges including high computational complexity and excessive parameters in crop disease recognition, hindering practical deployment. To address the above problems, we propose a lightweight full-depth separable and multi-scale aggregation network (FDS-MANet) to achieve high segmentation accuracy with reduced computational overhead. Specifically, a fully depthwise separable convolution architecture is proposed to reduce the number of parameters and improve running speed. In addition, we design a lightweight Atrous Spatial Pyramid Pooling with Attention Aggregation (ASPP-AA) module that can extract multi-scale features through depthwise separable atrous convolutions and attention mechanisms while reducing parameters and computational complexity. Then, we introduce the SimAM attention module to extract spatial and channel features, which compensates for the model's capability in handling high-frequency local features, thus further enhancing segmentation precision. Experimental results demonstrate that FDS-MANet achieves a mean intersection over Union (mIoU) of 79.93% with only 2.77M parameters on constructed multicategory disease dataset which outperforms competitive models such as DeepLabV3+ and PVTv2, etc, while reducing parameters by 93% and computational costs by 89%. As a result, the method we propose achieves a good balance between accuracy and efficiency, showing potential for real-time disease segmentation in real agricultural environments. • We propose full-depth separable and multi-scale aggregation network (FDS-MANet), a lightweight segmentation network for Panax notoginseng disease diagnosis. • We design the Lightweight Atrous Spatial Pyramid Pooling with Attention Aggregation module to enable efficient multi-scale feature extraction. • We introduce the Simple Attention Module (SimAM) to improve high-frequency detail capture and fine boundary segmentation without extra parameters. • Proposed FDS-MANet achieves 79.93% mIoU with only 2.77M parameters , surpassing PSPNet, Deeplabv3+, and PVTv2 while cutting 89–93% of computational costs. • The method demonstrates strong robustness in complex backgrounds, small lesions, and multi-disease regions , making it suitable for real-time agricultural applications.