Research On Pest Detection Based On Deep Learning

Agricultural pests and diseases have always been an important factor affecting the quality and yield of agricultural production.Therefore,automatic monitoring identification and timely control of pests in the process of engaging in agricultural production has a very high practical value in the field of agriculture.However,the large differences in the morphology and color of pests in the natural environment,coupled with the differences in natural conditions such as light and climate,make the accurate detection and identification of agricultural pests a great challenge.Traditional pest control in agriculture mainly relies on human observation to identify pests,which is time-consuming and laborious.Machine learning-based pest identification methods,which rely on manual extraction of features,require experts with experience and expertise to carry out,and machine learning methods are not robust and difficult to adapt to natural environments.With the rapid development of computer technology,computer vision technology based on deep learning has been applied and promoted in the fields of security and autonomous driving,and deep learning provides a new direction for pest detection.In this thesis,based on the target detection technology in the field of deep learning,we propose a target detection method based on attention mechanism and multi-scale feature fusion,and integrate the prediction results of multiple detection models through integrated learning ideas to pool ideas and further improve the detection accuracy of pests.The main work of this thesis includes:(1)We use a deep learning-based target detection algorithm to detect pests,and effectively improves the feature extraction capability of convolutional neural networks by introducing the ECA module into the feature extraction network to promote the network to focus more on the channels of interest and inhibit the network from extracting impurities and background features.Secondly,this thesis combines the convolutional neural network with the transformer encoder structure,which effectively improves the global feature processing ability of the model for pests.Meanwhile,this thesis proposes a cross-stage feature fusion strategy in the multi-scale feature fusion network and creates cross-stage branches on the feature pyramid network and path enhancement network,which effectively improves the feature representation of feature maps for small targets such as agricultural pests.(2)In order to further improve the accuracy of target detection methods for pest detection,we propose a multi-model fusion method for target detection based on integrated learning,by training multiple target detectors independently and using weighted frame fusion or stacking methods to fuse the prediction results of detectors to output,which makes the fused detection results improve in terms of detection accuracy.In this thesis,experiments are conducted on a large-scale very smallscale agricultural pest dataset,Pest24,and the experimental results illustrate the effectiveness of the above work.The experimental results illustrate that the attention mechanism gets a large improvement in detection accuracy at the cost of smaller parameters and operations,the transformer encoder plays an important role in extracting the global features of the image,and the cross-stage feature fusion strategy effectively fuses the contour information and semantic information of the low-level and high-level feature maps.Meanwhile,the target detection method based on the idea of integrated learning proposed in this thesis further improves the detection accuracy of pests through a collective decision-making approach.