| Agriculture and forestry are important basic industries in our country.They can integrate ecological benefits,economic benefits,and social benefits and play an important role in solving the problem of food shortages and improving farmers’ economic development.However,affected by factors such as climate change and crop layout adjustment strategies,the agriculture and forestry industries are often infested by a large number of pests,which undoubtedly becomes a major obstacle restricting the rapid development of agriculture and forestry.Although the use of a large number of pesticides has reduced the impact of pests to a certain extent,it will inevitably cause pollution and damage to soil,groundwater,and atmospheric environment.Part of the reason for non-scientific use of pesticides is that farmers do not have effective prevention and control information.In particular,it is difficult to accurately obtain information about the types of pests and the severity of pest outbreaks.Therefore,a sound pest monitoring system is of great significance to the sustainable development of agriculture and forestry.Among them,accurate identification of pests is the basis of pest monitoring.Traditional methods relying on manual identification and monitoring of pests have problems such as high labor intensity,low efficiency,and poor real-time performance.With the continuous development of computer image processing technology,in order to seek an efficient,objective and practical solution,this paper introduces the object detection technology of deep learning into the pest detection task,and launches the following research:(1)Through the field and the Internet collecting pest images in complex environments,a data set containing 16 common pests was constructed.It has a total of 6722 pictures,which were labeled and divided.In order to solve the over-fitting problem that may be caused by insufficient data in the subsequent training process,the number of pictures was expanded to 31997.This data set can be applied to the detection field related to pests.(2)A detector relying on cyclic attention was proposed based on the SSD one-stage detection algorithm.This algorithm combined Res Net-34 with the proposed cyclic attention block,while maintaining a high detection speed and powerful classification ability,it explored the factors that affect the poor detection effect of the detector on small objects,and achieved a good detection effect for small-sized objects and overlapping objects.This algorithm achieved 82.5% m AP on PASCAL VOC.In order to make the method achieve the best recognition effect on the pest data set in this paper,the scale and aspect ratio of the default boxes in the algorithm were adjusted,and the detection result reached88.7% m AP.(3)Because the most important thing in the pest detection task is the accuracy of recognition,in order to pursue higher recognition accuracy,it is improved on the basis of the two-stage detector Faster R-CNN.First,an efficient Vo VNet was selected as the backbone network.Second,to improve the defects of FPN,the same supervision signal was added to the features before fusion to narrow the semantic gap between features,so that the features can be merged more successfully.The top-level features whose information was lost due to downsampling were enhanced,and then the algorithm used the balanced pyramid to further improve the expressive ability of features.Finally,adaptive anchors were used to provide more accurate prediction results for the detected object.Experiments have proved the effectiveness of each component,reaching 83.2% m AP on the PASCAL VOC data set.The algorithm combined with the data augmentation strategy,reaching 92% m AP on the pest data set in this article. |
