| Deep learning is one of the key achievements in the field of artificial intelligence research,widely applied in the field of image processing.Deep learning based models require a large number of already annotated image datasets for training,but currently most datasets rely on manual annotation,which is costly and inefficient,and poses a time-consuming and labor-intensive problem.Therefore,the research on automatic image annotation methods and tool software is of great significance for promoting the development of deep learning model modeling.This paper introduces the theory of deep learning and the classification of image annotation technology,studies the modeling method of image annotation model based on rice disease spot image,and further develops the automatic image annotation system based on agricultural images relying on small sample learning method.The specific work is as follows:(1)Image dataset partitioning and preprocessing.This article divides the agricultural image dataset into training and testing sets.At the same time,different data enhancement methods are used to preprocess the images,including research on various image enhancement methods such as flipping,rotating,cropping,deformation,scaling,and image normalization methods.This provides a large amount of image data for model training.Research has shown that the partitioning ratio of the dataset,data augmentation,and image normalization methods can accelerate model training speed and enhance the generalization ability of the model.(2)Studied the improved model structure based on YOLOv5.In order to enhance the deep learning model’s ability to annotate lesion images under small sample conditions,this paper improves the YOLOv5 model structure to YOLOv5-TR-Bi FPN model structure.Bi FPN and Vi T are integrated into YOLOv5-TR-Bi FPN structure,which enhances the Receptive field of YOLOv5-TR-Bi FPN model and the ability to accurately locate the target.CIOU is used to calculate Loss and NMS to improve the accuracy of the target frame.The experimental results show that the YOLOv5-TR-Bi FPN model has an average accuracy of 73%,which is 3.8% higher than YOLOv5 s.It also achieves good annotation performance for small and dense disease spots in rice disease spot images.Tested on 30,10,and 5 images of rice blast disease,the YOLOv5-TR-Bi FPN model achieved a m AP of 89.6% on a dataset of 30 images and quickly converged,proving that the YOLOv5-TR-Bi FPN model can quickly annotate disease spot images in small samples.(3)We have developed an automatic annotation system based on agricultural images.In order to solve the current problem of lacking a professional software that can realize automatic annotation of agricultural automatic image annotation,this study designed an automatic annotation system based on agricultural images using the development framework of Python+Py Qt5+Flask based on the automatic annotation model of rice disease spot images,which realized the development of graphical interfaces and server interfaces,and completed the needs of interaction and cooperation between client and server.The previous backend integration approach achieved training of image annotation models and automatic annotation of a large number of images.The graphical interface operation steps of this system are simple and intuitive,improving the time-consuming and labor-intensive disadvantage of manually labeling images,and achieving automatic labeling for agricultural images.Using grape black rot image testing and statistics,the labeling accuracy reached 97.73%,the omission rate was 2.14%,and the error rate was 0.13%,proving that the system can accurately and automatically label agricultural images in small sample sizes. |
