WE-DeepLabV3+: A lightweight segmentation model for Panax notoginseng leaf diseases

Abstract

Panax notoginseng plays an important role in traditional Chinese medicine. However, diseases pose a significant threat to the quality and yield of P. notoginseng. The main challenge related to the identification of P. notoginseng leaf diseases is how to achieve good performance in the case of small diseased spots on P. notoginseng leaf, overlapping edges of diseased leaf and the difficulty in mobile deployment. A lightweight semantic segmentation model Window Efficient-DeepLabv3+ is proposed for segmentation and quantification of P. notoginseng leaf diseases. We propose the Window Attention-ASPP module and present a hierarchical stacking of features, which improves model accuracy for minor target lesions while reducing parameters. In addition, a lightweight backbone network MobileNetV2 is utilized as a feature extraction module. The decoding stage introduces the Efficient Channel Attention module, which effectively improves the accuracy of the segmentation of the blade contour. Experimental results yielded the Mean Intersection Over Union, Mean Precision, and Mean Recall metrics of WE-DeepLabV3+ network to be 82.0 %, 87.6 %, and 92.4 % respectively, outperforming other segmentation models such as UNet, PSPNet, CaraNet, SegNet, and BiSeNetV2. Moreover, the number of parameters has been reduced by 90.6 % with only 5.1 M parameters. Finally, the method is used to quantify the disease of P. notoginseng leaf, the error is only 1.15 % and 0.82 %, which proves that it can quantify the disease severity accurately. Thus, the proposed method holds great significance for raising the yield and quality of P. notoginseng, also providing reliable guidance for precise fertilization and drug control.