Research On Detection And Recognition Of Crop Pests And Diseases Based On Similar Modalities

Real-time,effective,and accurate detection of agricultural pests and diseases is crucial in ensuring agriculture’s healthy and sustainable development.In recent years,with the gradual deepening of climate warming,agricultural fertilizer accumulation,environmental degradation,and other unfavorable factors,the development of pests and diseases has shown multiple,frequent,and severe situations.The causes of disasters are gradually becoming more complex,so it is urgent to research and analyze the causes of pests and diseases and carry out precise prevention and control in a targeted manner.With the rapid development of artificial intelligence technology,researchers will be based on deep learning vision technology widely used in automatically detecting and identifying crop pests and diseases.The latest research results show that deep network learning-based models are adequate for general pest and disease target detection.Still,using the available detection methods for pest and disease problems with similar modalities,such as appearance and shape,color and texture,size distribution,and background interference,is difficult.Therefore,how to effectively improve the characterization of similar-modal pest and disease target features is a vital issue that needs to be solved.Based on this,this dissertation mainly focuses on the similar modalities such as shape,texture,size,and background of similar pests.It carries out essential algorithm research on the problems of insufficient discriminative feature fusion of similar pests,inadequate long-distance semantic characterization of similar pests,and incompatibility of similar pests and pathogens in symbiosis,etc.The specific contents and innovations are as follows:1.Aiming at the problem that pests of the same family and genus are highly similar in shape,texture,and scale modality,densely distributed,and challenging to accurately identify,a domain adaptive detection algorithm for similar pests is proposed.Firstly,a two-channel attention mechanism is designed in the backbone network to enhance the extraction of discriminative features of similar modal targets.Then,a non-local cooperative work model is incorporated to extract the saliency detail information of the discriminative features for fusion processing.Next,the problem of under-representing similar pest images due to minor scale differences and dense distribution is solved by fusing scale-jumping calibrated convolution.Finally,the detection network without an anchor frame is constructed by mixing centroid control loss through the adaptive perceptual head network.Significant feature extraction and activation are completed for smallscale similar pests with dense samples to achieve intelligent recognition and localization of highly similar pest images in appearance.Sufficient experiments are conducted on the dataset PestNet-AS.The results show that the anchor-less frame single-stage target detection algorithm proposed in this dissertation has optimal detection accuracy compared to the mainstream target detection algorithms in the same period.Meanwhile,the model solves the balanced requirements of accuracy and efficiency and improves the model detection speed while ensuring optimal accuracy.2.Aiming at the problem that the scales of similar diseases of multiple crops are highly similar in shape,color,and background modes,and the diseases have large-scale changes and severe overlap,a network structure for the integration of adjustable cross windows and feature pyramids for multi-scale similar diseases is proposed.Firstly,the algorithm constructs a cross-cross Transformer backbone network,which filters the background interference information of similar disease modalities through the crosscross interlaced growth pattern and improves the extraction of significant features in similar modalities.Meanwhile,the cross-crossing window is conducive to reducing the computation of the model.Secondly,the multi-head attention mechanism is reconstructed in the high-level semantics,and the interactive spatial multi-head attention mechanism formed can effectively extract the significant discriminative features of similar disease modalities and enhance the a priori guidance ability of the network.Finally,a global association enhancement module is designed in the neck feature pyramid network to solve the problem of local feature confusion caused by simply superimposing the information of low-and high-level feature maps.With missing datasets for similar diseases,this dissertation forms a multi-category similar disease dataset,MCD2022,and fully validates the proposed method on this dataset.The experimental results show that our proposed method has significantly improved the characterization ability of similar disease modalities and the interpretability of the network,achieving the best results of the contemporaneous detection methods.3.Aiming at the problem that the existing detection algorithms generally complete the network design in the task of a single dataset of similar pests or diseases but fail to deeply mine the unified representation method of feature domain information with homogeneous problems,a disease and pest detection method oriented to similar modes is proposed,which mainly includes redefining the classification loss,bounding box loss,and reconstructing the content of three parts:sample selection method.Firstly,the objective function of the classification loss adjustable with additive angle cosine is used to increase the penalty for the salient features of the high similarity modal samples so that the model can converge to the optimal result quickly and accurately.Then,the detection frame is analyzed,and a localization loss function is constructed suitable for similar modes of pests and diseases,named SMIoU.Then,according to the characteristics of personality and commonality of similar modes,a sample learning selection strategy is constructed,which can dynamically adjust the significant modal differences of similar samples and automatically adjust the thresholds of different samples according to statistical attributes to help the model select relatively high-quality training samples to complete the prediction process.A large number of experimental results on the constructed similar modal pest dataset(SMCPD)show that the strategy also strengthens the training effect of the SMIoU loss function,improves the problem of target boundary classification fuzziness,and improves the interclass separability of cross-similar modalities,to ensure the detection effect of similar pest targets.Therefore,the target loss function reconstructed in this dissertation and the sample learning strategy adopted in this dissertation can effectively alleviate the problem of incompatibility of similar modes of pests and diseases and lay a theoretical foundation for exploring the target detection methods of similar modes.