层次特征与语义融合解码的苹果叶片病害语义分割

Hierarchical Feature and Semantic Fusion Decoding for Semantic Segmentation of Apple Leaf Disease

摘要:

苹果叶片病害语义分割是智能农业病害监测的重要任务。针对自然场景中病害分割病斑尺度变化大、边界模糊、复杂背景干扰的挑战,本文提出层次特征与语义融合解码的苹果叶片病害语义分割模型。首先,提出层次特征网络(Hierarchical feature network,HFN),通过五阶段特征提取和残差结构实现从边缘轮廓到全局语义的建模。其次,为增强对病斑尺度变化和边缘结构的适应能力,设计边缘采样空洞空间金字塔模块(Edge sampling atrous spatial pyramid,ES-ASPP),捕捉多尺度语义信息并细化边缘结构。最后,由于语义分割任务中需要高效解码器以还原空间信息,因此提出语义融合解码器(Semantic fusion decoder,SFD),协同表达浅层细节与深层语义,提升边界还原与小病斑检测能力。试验结果表明,本文所提方法在苹果叶片病害数据集上的mIoU较基线模型提升5.7个百分点,参数量下降6.2%;且在PlantSeg数据集泛化试验中相较DeepLabV3Plus mIoU提升4.3个百分点,具备良好的分割性能,适用于自然场景中农业病害分割任务。

Abstract:

Semantic segmentation of apple leaf disease is an important task for disease monitoring in smart agriculture. Aiming at the challenges of large scale changes, blurred boundaries, and complex background interference of disease segmentation spots in natural scenes, a semantic segmentation model for apple leaf disease with hierarchical feature and semantic fusion decoding was proposed. Firstly, hierarchical feature network (HFN) was proposed to achieve modelling from edge contour to global semantics through five-stage feature extraction and residual structure. Secondly, in order to enhance the adaptability to spot scale variations and edge structure, the edge sampling atrous spatial pyramid (ES-ASPP) module was designed to capture multi-scale semantic information and refine the edge structure. Finally, since an efficient decoder was required in the semantic segmentation task to restore spatial information, semantic fusion decoder (SFD) was proposed to synergistically express the shallow details and deep semantics to enhance the boundary restoration and small lesion detection. The experimental results showed that the mIoU of the proposed method on the apple leaf disease dataset was improved by 5.7 percentage points compared with that of the baseline model, and the number of parameters was decreased by 6.2%;and the mIoU was improved by 4.3 percentage points in the generalisation experiment on the PlantSeg dataset, which can provide a good segmentation performance, and it was suitable for the task of segmentation of agricultural diseases in natural scenes.