Research On Microscopic Image Detection Algorithm For Wheat Scab Spores

Wheat scab is a destructive plant disease that has caused significant damage to wheat crops worldwide.The detection of wheat scab spores is essential to ensure the safety of wheat production.However,traditional detection methods require expert opinion in their detection processes,leading to less efficiency and higher cost.Currently,deep learning-based fungal spore detection methods have provided references for the detection of wheat scab spores.However,these methods still need further improvement in detection accuracy,and the large number of parameters and computations make the model structures complex.Moreover,due to the scarcity of wheat scab spore microscopic image data,it is difficult to obtain large-scale and accurately labeled images.Therefore,to provide strong support and assistance for the prevention and control of the wheat industry,this dissertation aims to conduct research on microscopic image detection algorithm for wheat scab spores.The main research contents and contributions are as follows:(1)Spore detection based on fully-supervised learning.This dissertation proposes a spore detection method,SporeDet,based on a holistic architecture called‘backboneFPN-head’.Specifically,the method utilizes RepGhost with FPN to fuse feature information from the backbone while minimizing the model’s parameters and computation.Additionally,a task-decomposition channel attention head(TDAHead)is designed to predict the classification and localization of FPN features separately,thereby improving the accuracy of spore detection.Furthermore,a feature reconstruction loss(RecLoss)is introduced to further learn the features of RGB images during the training process,which accelerates the convergence of the model.The proposed method is evaluated on spore detection datasets collected from the Anhui Academy of Agricultural Sciences.The experimental results demonstrate that the SporeDet method achieves a mean average precision(mAP)of 88%,and the inference time of the model reaches 4.6 ms on a 24GB GTX3090 GPU.Therefore,the proposed method can effectively improve spore detection accuracy and provide a reference for detecting fungal spores.(2)Spore detection based on semi-supervised learning.This dissertation proposes a spore detection method,Semi SporeDet,based on the teacher-student model mutual learning mechanism.Specifically,the method employs the Burn-In strategy to perform joint training on labeled and unlabeled data.Moreover,during the pseudo-label generation phase,a categorization scheme is implemented to differentiate between reliable,candidate,and unreliable pseudo-labels,aiming to select more accurate pseudo-labels.A dynamic adaptive threshold is introduced to dynamically calculate a high threshold for filtering out pseudo-labels,thereby generating high-quality pseudo-labels during the semi-supervised spore detection training.Additionally,a domain-aware loss is introduced to optimize the model’s features and differentiate between labeled and unlabeled data,thereby improving the model’s generalization performance.The experimental results demonstrate that the SemiSporeDet method outperforms the baseline method,achieving an mAP of 83.8%in a fully labeled setting and 74.1%and 77%in partially labeled settings with labeling rates of 5%and 10%,respectively.Therefore,the introduction of semi-supervised learning effectively improves the training performance of labeled data.In summary,the above algorithms can provide robust support and assistance for the prevention and control work of the wheat industry,and offer new ideas and methods for the detection and diagnosis of other other fungal spores.