IntroductionPrecise identification of crop insects is a crucial aspect of intelligent plant protection. Recently, with the development of deep learning methods, the efficiency of insect recognition has been significantly improved. However, the recognition rate of existing models for small insect targets is still insufficient for insect early warning or precise variable pesticide application. Small insects occupy less pixel information on the image, making it more difficult for the model to extract feature information.MethodsTo improve the identification accuracy of small insect targets, in this paper, we proposed S-ResNet, a model improved from the ResNet, by varying its convolution kernel. The branch of the residual structure was added and the Feature Multiplexing Module (FMM) was illustrated. Therefore, the feature expression capacity of the model was improved using feature information of different scales. Meanwhile, the Adjacent Elimination Module (AEM) was furtherly employed to eliminate the useless information in the extracted features of the model.ResultsThe training and validation results showed that the improved residual structure improved the feature extraction ability of small insect targets compared to the original model. With compare of 18, 30, or 50 layers, the S-ResNet enhanced the identification accuracy of small insect targets by 7% than that on the ResNet model with same layer depth.

Weed suppression is an important factor affecting crop yields. Precise identification of weed species will contribute to automatic weeding by applying proper herbicides, hoeing position determination, and hoeing depth to specific plants as well as reducing crop injury. However, the lack of datasets of weeds in the field has limited the application of deep learning techniques in weed management. In this paper, it presented a dataset of weeds in fields, Weed25, which contained 14,035 images of 25 different weed species. Both monocot and dicot weed image resources were included in this dataset. Meanwhile, weed images at different growth stages were also recorded. Several common deep learning detection models—YOLOv3, YOLOv5, and Faster R-CNN—were applied for weed identification model training using this dataset. The results showed that the average accuracy of detection under the same training parameters were 91.8%, 92.4%, and 92.15% respectively. It presented that Weed25 could be a potential effective training resource for further development of in-field real-time weed identification models. The dataset is available at https://pan.baidu.com/s/1rnUoDm7IxxmX1n1LmtXNXw; the password is rn5h.