Research On Accurate Segmentation Algorithm Of Green Target Fruit In Orchard Environment

With the continuous innovation of machine vision technology in the fields of agricultural equipment and smart farm,the basic and core technologies of agriculture are also rapidly tamped and broken through,promoting the development of agriculture towards automation and intelligence.Among them,the accurate and efficient identification of target fruits plays an important role in intelligent picking,growth monitoring,yield estimation and other applications,and it is also the basic guarantee to realize these agricultural automation applications.However,in the natural orchard environment,unstructured scene,unconventional fruit posture,multi illumination environment,multi-angle position acquisition,homochromatic background and other interference factors restrict the fruit recognition effect.In this study,the green fruit with greater difficulty in recognition is selected as the research object,for improving the green fruit recognition effect in the natural orchard environment from multiple dimensions.Then drive the agricultural machinery and equipment to improve the picking success rate and reduce the yield estimation error.The specific research contents are as follows:(1)Construct the data sets of green fruit for detection and segmentation.In order to make the model have the ability to accurately identify fruits in the real environment,the collected images should include complex and diverse orchard scenes as complete as possible to supervise the training convergence of the subsequent proposed model.Therefore,through five specific steps of image acquisition,cleaning,labeling,data set generation and statistical analysis,two green fruit data sets of green apple and immature persimmon are constructed,which support two task types of object detection and case segmentation,and support the follow-up research and experiment of this thesis.(2)A fruit instance segmentation model named RS Net(Robust Segmentation Network)is proposed.Firstly,aiming at improving the accuracy and robustness of fruit recognition and the segmentation problem of green apple image in natural orchard environment,based on the instance segmentation model Mask RCNN,and fully considering many interference factors in natural scene,RS Net is obtained through optimization to strengthen the understanding ability of operation scene and improve the accuracy and robustness of model segmentation.The balance feature pyramid(BFP)module is embedded in the model to eliminate the semantic dilution problem caused by the top-down fusion of non adjacent resolution features in the feature pyramid network(FPN),and aggregate the context feature information within the whole image through the Gaussian nonlocal attention mechanism in BFP to alleviate the homochromatic background,branch occlusion The segmentation interference caused by fruit overlap and other factors on the fruit with missing characteristics.(3)A fruit instance segmentation model named Fovea Mask is proposed.In addition to the metrics of segmentation accuracy,embedded mobile agricultural machinery equipment needs to meet the indexes of multiple dimensions such as model detection speed,capacity,complexity and generalization ability.Therefore,an anchor free fruit segmentation framework Fovea Mask is proposed.Firstly,the model classifies and regresses each spatial position on the feature map directly in the way of full convolution,and then realizes the pixel level classification of each fruit instance through the embedded mask branch.Similarly,in order to improve the segmentation robustness of the model in the orchard environment,the model embeds the spatial attention module in the mask branch.In addition,the Mask Io U branch is introduced at the end of Fovea Mask to obtain the corrected fruit mask confidence.The whole network architecture does not involve the design of anchor frame and related operations,which greatly improves the generalization ability of the model,alleviates the computing and storage resources,and better balances the relationship between accuracy and efficiency.The above research strengthens and improves the scene understanding ability and segmentation accuracy of the model in the orchard environment,provides theoretical and technical support for intelligent picking and orchard yield estimation,drives the development of agricultural intelligent application in the direction of practicality,and can also further provide theoretical reference for the identification of other fruits and vegetables.