关键词: 叶片病害, CBAM, 双向特征融合, Mask R-CNN, NMS

Abstract: When the traditional Mask R-CNN network detects the target, feature loss and feature confusion will occur, and for the dense small target, it is easy to miss detection, false detection and other problems. In order to solve this problem, this paper proposes a leaf disease detection method combining attention mechanism and bidirectional feature fusion. Firstly, two kinds of artificial noises such as Gaussian noise and salt and pepper noise, were added to the leaf picture during the construction of the data set to imitate the complex noises in nature and improve the diversity of data. Secondly, combining the PAFPN structure with the CBAM attention mechanism, the CBAM-PAFPN structure is generated to replace the FPN structure of Mask R-CNN network and optimize the feature extraction mode of Mask R-CNN network. Finally, replace the original NMS filtering candidate box with Soft-NMS. The experimental results show that for the noiseless data set, mAP value increases by 0.46% and Recall value increases by 2.24%. The average error detection rate is 1.34%, a decrease of 3.28%, about 1/4 of the original network, the average missing detection rate is 0.12%, a decrease of 2.19%, about 1/20 of the original network. The improved network has increased the accuracy of detection and positioning, which provides technical support for the effective detection of leaf diseases of different sizes and densities.

Key words: leaf disease, CBAM, bidirectional feature fusion, Mask R-CNN, NMS