Fruit Flower Segmentation Method Based On Semantic Segmentation Network And Fuzzy Energy Active Contour Model

In the process of fruit production,the accurate estimation of flowering intensity of fruit flowers is a key factor affecting growers to make crop management decisions,which is closely related to fruit yield.Using computer vision to estimate flowering intensity is a trend direction.The existing computer vision methods are mainly for fruit and flower image segmentation,and then estimate flowering intensity.Although the current computer vision flower segmentation methods can achieve better segmentation accuracy,these methods have limitations.Most of them can only be used under specific conditions.When the flower species,lighting conditions,weather conditions,the relative position of the camera and fruit trees change,these methods either fail completely or need to adjust the parameters,the accuracy and practicability of the method are reduced.This thesis presents a method of automatic segmentation of fruit and flower images in different environments,which can segment many kinds of fruit and flower images including apple flower,peach flower and pear flower.This method uses the fine tuned Deep Lab-Res Net network and trains only on the apple flower data set.It can be applied to the automatic detection of a variety of fruits and flowers,mainly to identify the flower area.In order to make up for the lack of positioning accuracy caused by the introduction of several pooling layers and lower sampling layers in the Deep Lab-Res Net network,two fuzzy energy active contour models with shape priors are proposed in this thesis.The flower recognition results of the network are regarded as shape priors,and the color information of the original image is combined to obtain more accurate segmentation results.The main work of this thesis is as follows:（1）This thesis presents a new fruit and flower image segmentation model for outdoor fruit and flower images with unlimited weather,unlimited illumination,diverse species and complex background.The model has three typical characteristics:a)it is an automatic image segmentation system;b)It is robust to complex environment and different illumination conditions;c)It has good generalization ability for fruit flower images of different varieties and colors,and can be applied to fruit flower image segmentation of multiple species.（2）In this thesis,two improved fuzzy energy active contour models with shape priors are proposed as the post-processing of convolutional neural network.The two fuzzy models proposed in this paper effectively improve the rough results of semantic segmentation obtained by convolutional neural network through the minimization iterative process of energy functional.（3）The two methods and three typical flower segmentation methods are compared on four public apple flower,peach flower and pear flower datasets.The quantitative results show that the average F₁score of pixel level of this method on four datasets reaches 82.0%and 83.1%,which is higher than the latest method on the same dataset.At the same time,the calculation cost of this method is relatively low.