MFNet: Multi-branch feature-aware network for multi-class disease segmentation of Panax notoginseng

Panax notoginseng (Burk.) F. H. Chen, belonging to Araliaceae, mainly cultivated in Yunnan Province, China, accounting for over 90 % of the nation’s total yield. However, the growth of Panax notoginseng is stressed by various diseases, thereby severely reducing the yield and quality. Therefore, accurate and automatic diseases segmentation is essential for the prevention and control of Panax notoginseng diseases. Due to the visual similarity, variations in scale and shape, and edge complexity among different diseases, existing disease segmentation approaches is challenging in the extraction of high-frequency local information and cross-layer information exchange, leading to low segmentation performance. To address the above problems, the multi-branch feature-aware network (MFNet) is proposed for Panax notoginseng leaf multi-class disease segmentation, which combines three parallel branches to fuse multi-level features while integrating low-frequency global and high-frequency local features. Specifically, the efficient convolutional gated channel Attention (EC-GCA) module is proposed for capturing high-frequency local information from small disease spots. Additionally, the efficient depthwise separable atrous spatial pyramid pooling (EDS-ASPP) module is developed to enhance multi-scale disease perception capabilities by utilizing multi-scale separable atrous convolutions and cross-channel interactions. Finally, the gather and distribute mechanism is introduced to realize the effective fusion of global features and local features. The experimental results show that the proposed method achieves the competitive segmentation performance, with 81.94 % mIoU and 89.51 % mDice on the Panax notoginseng disease dataset. Thus, the proposed model can provide theoretical support for the precise quantification of diseases. • A multi-branch feature-aware network is proposed for multi-class Panax notoginseng leaf diseases. • The gather-and-distribute mechanism is introduced to fuse multi-level information effectively. • The convolutional gated channel attention module is proposed to capture high-frequency local information. • The depthwise separable atrous spatial pyramid pooling module is proposed to enhance multi-scale feature perception. • A multi-class Panax notoginseng disease dataset is constructed in a complex environment.