一种基于改进MaskRCNN的纸病诊断算法

An Improved MaskRCNN Based Paper Disease Diagnosis Algorithm

关键词:  纸病诊断  MaskRCNN  VOVNet  PrRoIPooling  SPANet

Key Words:paper disease diagnosis  MaskRCNN  VOVNet  PrRoIPooling  SPANet

摘要:本研究提出了一种基于改进MaskRCNN网络的纸病诊断算法。该算法首先在原有的MaskRCNN网络的基础上，使用轻量化头部骨干网络VOVNet和精细化的RoIPooling（PrRoIPooling）对原网络模型进行改进，以减少原网络模型的参数使用量，提升检测分类速度；其次添加空间金字塔注意力机制（SPANet），以解决原网络模型对于小目标检测精确度不高的问题。采集4 000多张纸病图像对本研究提出的算法进行仿真验证。结果表明，改进的MaskRCNN模型比原网络模型在平均精度上提升了3个百分点，速度上提升了15%，能够满足纸病诊断的实时性和准确性的实际需求。

Abstract:This paper proposed a paper disease diagnosis algorithm based on an improved MaskRCNN network. Firstly， this algorithm improved the network model by using a lightweight head backbone network VOVNet and a Precise RoIPooling （PrRoIPooling） on the basis of the original MaskRCNN network， in order to reduce the parameter usage of the original network model and improve the detection and classification speed. Secondly， a spatial pyramid attention mechanism （SPANet） was added to address the issue of low accuracy in detecting small objects in the original network model. More than 4 000 paper disease images were collected for simulation verification of the proposed algorithm. The results showed that the improved MaskRCNN model had increased average accuracy by 3 percentage points and speed by 15% compared to the original network model， which could meet the practical requirements of real-time and accuracy in paper disease diagnosis.