关键词: 叶片病虫害识别, 多层级特征编码, 多尺度特征融合, 通道注意力, 特征表达

Abstract: Aiming at the problem that traditional detection networks are difficult to extract the feature information of crop leaf pest and disease accurately and efficiently, a multilevel and multiscale feature fusion recognition algorithm for crop leaf pest is proposed through the improvement of YOLOv8 network. Firstly, a multilevel feature coding module is constructed to learn the comprehensive feature representation by learning the direct feature relationships of different levels of features. Then, a multiscale spacechannel attention module is designed on the basis of Transformer to capture the complementary relationships between different scales of features by learning the comprehensive multiscale feature representation patterns such as finegrained and coarsegrained, and all feature representations are effectively. The whole feature representations are fused, and the better recognition results are obtained.Finally, the experimental validation is conducted on the Plant Village public dataset, and the results show that the proposed improved method can effectively improve the alignment accuracy and accurately recognize different pests and diseases existing on the leaves of crops at the same time, and the mAP 0.5 for tomato leaves detection reaches 88.74%, which is 8.53% higher than the traditional YOLOv8 method, without significant increase in computation time. The ablation experiments also fully demonstrate the effectiveness of the proposed modules, which can better achieve highprecision leaf insect and disease recognition and provide a strong support and guarantee for the intelligent management of farmland.

Key words: recognition of leaf disease and pest, multilevel feature coding, multiscale feature fusion, channel attention, feature expression