Research On Small Object Detection For Early Apple Leaf Diseases

Apple leaf diseases have significant impacts on apple quality and productivity.Thus,the implementation of accurate disease detection and corresponding control measures in the early stages can effectively improve apple quality and provide a strong guarantee for the highquality development of the apple industry.However,early apple leaf disease often represents very small size disease spots.So it is difficult to extract the disease spot features and locate the disease spots when using the deep learning-based network to detect the disease,which makes the early disease detection result not ideal.Thus,this paper conducts research from the following aspects:(1)Focusing on the difficulty of small spot feature extraction,a novel detection model called Representation-Enhanced RCNN(RE-RCNN)is proposed to perform accurate detection of early apple leaf disease spots.Firstly,an object-enhanced branch is proposed to achieve feature enhancement of small disease spots by introducing a small disease spots feature enrichment extractor(SDSFEE).Secondly,a SCMLoss is proposed to balance the inter-class differences of various size disease spots under the same category.Thirdly,an one2 one computation strategy is leveraged to sample data reasonably during the training process.From the final experimental results,it can be seen that the proposed model could achieve outstanding performance on the early apple leaf disease detection task.(2)Focus on the difficulty of locating small spots,a novel IOU-based regression loss function Relation-IOU Loss is proposed.It is used to guide the overall parameter update direction to further improve the detection model localization accuracy in the small spot detection task.The function quantifies the differences between the disease spots region proposals and the disease spots ground truth by calculating Intersection over Union,centroid Euclidean distance,and height-width difference,from the geometrical perspective combined with the model accuracy calculation method.To address the problem of gradient disappearance and gradient explosion in the regression process,it combines the advantages of existing regression loss functions and uses the exponential function to optimize the calculation of the bounding boxes differences.The experimental results show that the Relation-IOU Loss can effectively assist the model for the early apple leaf disease detection task.(3)The early apple leaf disease detection system with the Django framework is developed based on this paper’s research and the actual production needs.It implements functions such as the accurate detection of early common apple leaf diseases,the description of the measured diseases,and the corresponding disease-control methods.It provides reliable technical guarantees for disease control management and precise application.This helps to reduce the loss of resources such as labor and economic costs.