Superpixel Algorithm Improvement And Research On Arabidopsis Thaliana Leaf Segmentation Based On Deep Learning

Superpixel algorithm is an important preprocessing tool for computer vision.In recent years,it has been widely used in various fields of computer vision,especially in plant leaf segmentation and farmland remote sensing image analysis.The advantages of superpixel algorithm include: it can reduce the influence of noise and redundant information on subsequent use,preserve the structural characteristics of the image,reduce the number of image elements,and make it possible to perform large-scale calculations on the image.In practical applications,the classic superpixel algorithm can’t show good boundary accuracy when the number of superpixels is small,which brings great difficulties to the further image analysis.In order to make up the shortcoming,a superpixel boundary optimization framework aimed at improving the boundary accuracy of the classic superpixel algorithm is proposed.The proposed framework mainly consists of three parts: First,based on the proposed information measurement function,the under-segmented superpixels generated by the classic superpixel algorithm are screened out.Secondly,the under-segmented superpixels are subtly segmented by using two invariant clustering centers to improve the accuracy of the superpixel boundary.Finally,the number of superpixels is reduced by merging the smaller superpixels generated in subtly segmentation,so that the number of superpixel boundary optimization results remains the same as that of the initial superpixels.The superpixel boundary optimization framework has been tested using the BSDS500 dataset.The qualitative and quantitative evaluation show that the performance of the classic superpixel algorithm has been improved after using the superpixel boundary optimization framework.Especially in the case of a small number of superpixels,the effect is more significant.In the task of automatic monitoring of plant growth,changes in the state of plant leaves can reflect the status of plant growth.Therefore,it is very meaningful to realize the automatic segmentation of plant leaves.As a typical model plant,Arabidopsis thaliana has the characteristics of fast growth and widespread distribution in our country.Therefore,the selection of Arabidopsis thaliana as the research target and the study of plant leaf automatic segmentation have representative significance,and provide reference for the research of other plant leaf automatic segmentation.In recent years,deep learning models have been widely used in image segmentation tasks.Deep learning models have many advantages,such as high portability and high accuracy.Among the deep learning models,U-Net shows its powerful segmentation ability.It has the characteristics of simple model,high segmentation accuracy,and suitable for small datasets.Therefore,this study chooses the improved U-Net model to train and test on the task of Arabidopsis thaliana segmentation to determine the effectiveness of the deep learning model in the task of Arabidopsis thaliana leaf segmentation.The main work of this paper includes the following aspects:(1)The development of current superpixel algorithms is summarized,and the characteristics and shortcomings of the existing superpixel algorithms are analyzed.The reasons for the deficiency of the classical superpixel algorithm and the characteristics of the typical under-segmented superpixel are deeply analyzed.(2)An optimization framework for under-segmented superpixel boundary is introduced in detail.First,the information measurement functions used to screened out under-superpixels are introduced.Then,the module for screening under-segmented superpixels is introduced.Next,the under-segmented superpixels are subtly segmented.Finally,the segmentation results are merged according to the similarity in order to keep the same as the initial number of superpixels.(3)The basic principles of deep learning and the development of deep learning in recent years are introduced.The improved U-NET model is used to train and test the Arabidopsis thaliana leaf segmentation task,and the effectiveness of deep learning model in Arabidopsis thaliana leaf segmentation task was determined.