A novel farmland pest identification method based on transfer learning and improved YOLOv7

Abstract

ABSTRACT A novel farmland pest identification method based on transfer learning and improved YOLOv7 was proposed in this research, which aims to adapt to the practical application state of farmland pest identification. In this research, two datasets that are entirely different from the original dataset were built based on the original dataset with complex backgrounds. To create the improved model, the attention mechanism module was additionally incorporated into the network. Meanwhile, the improved model was applied to the original dataset following two-time transfer learning on the aforementioned two datasets. Finally, the model with the SimAm attention mechanism module achieves the best mean average precision of 72.88% on the test set, which is 2.72%, 6.34% higher than that of the model with the CA attention module, and the model with the ECA attention module, as can be seen from the experiment results. Furthermore, the three improved models may reach a mean average precision of 78.91%, 75.96%, and 72.04% on the test set, respectively, if two types of small object pests, No. 9 and No. 10, were deleted. Comparing the suggested method’s mean average precision with other methods that use the same dataset increases it by a maximum of 63.36% points. It is, therefore, demonstrated that the method suggested in this research can significantly improve the overall identification precision for agriculture pests in datasets with complicated backgrounds.