Salient Object Detection Based Study On Apple Leaf Disease Classification And Fruit Segmentation

With the continuous improvement of sensors and actuators,computer vision is gradually being used in fruit picking,crop control,harvesting and many other tasks,accelerating the pace of agricultural mechanization and intelligence.Among these,vision systems that enable accurate classification of plant disease types and precise positioning and segmentation of fruit fruits are of paramount importance to the various intelligent devices used in smart agriculture.However,many of today’s advanced models use data sets that are collected under controlled conditions such as indoors.In practice,due to the complexities of changing light conditions,overlapping and shading of leaves and fruits,dense distribution and color differences in real orchards,these models often result in misclassification,omission and mixing of machines,which seriously affects the normal work of intelligent devices and reduces their efficiency.The classification of apple leaf disease types and the segmentation of green fruits is explored from different perspectives and their effects are enhanced,making a contribution to agricultural modernization and informatization,as follows:(1)Construction of green fruit classification,detection and segmentation dataset.In order to make the model work more efficiently in the actual orchard environment,this work focused on the complexity and diversity of orchards when collecting orchard images,and collected a variety of real scenes of orchards in different states.To this end,this work has divided the acquisition dataset into four steps: image acquisition and cleaning,image dataset annotation,inspection and statistics of annotated images,and dataset generation.This work has constructed the dataset of apple fruits and leaves and peach fruits by the above steps,and supported two types of tasks: apple leaf and fruit disease classification and apple fruit instance segmentation.(2)The proposed saliency object detection model DEANet(Deep Enhanced Attention Network).It mainly consists of two parts: the dimensional-channel-spatial fusion(CSF)block and the edge-optimized loss module.The proposed CSF module uses an attention mechanism to capture the spatial dependence and channel dependence between RGB images and depth images,which can better exploit the complementary information between the two modal images and ensure the thoroughness and completeness of extracting useful information features.In addition,this work proposes an edge-optimized loss function to obtain smoother salient object edges by supervising the edges of the objects.(3)A two-stage network CBNet(Cascade Backbone Network)based on salient object detection and cascade backbone network is proposed for the classification of apple leaf diseases in the field.This work use Mobile Vi T partially instead of CNN for feature extraction.To better extract features from images,this work proposed a FR(Feature Refinement)module,which contains a dimensional attention mechanism to mine the connections between feature channels by dimensional attention,so that meaningful feature vectors in images can be extracted more effectively.This work also proposed a novel cascaded backbone network to replace the traditional decoder structure,which helps us to fuse the extracted feature information more effectively.(4)Proposed a two-stage fruit segmentation model Edge Seg Net based on salient object detection and boundary guidance.This work proposed a boundary-guided fruit segmentation framework based on a complex environment for improving the segmentation accuracy of obscured or overlapping fruits by accurately extracting the edge features of fruits and edge refinement of the fruit locations with accurate edge features to achieve accurate fruit segmentation.The proposed Global Localization Module(GLM)can accurately localize potential target regions using highlevel semantic information and local detail information.This work proposed the Multi-Scale Localization Block(MSLB)to accurately localize targets within a potential target region using rich multi-scale features.This work also proposed the Boundary-Aware Module(BAM),which makes full use of the edge information in the edge images to refine the target boundaries and make the boundaries smoother.In summary,this work has conducted a large number of experiments on the constructed apple leaf and fruit datasets,and the results of these experiments show that the above research improves the accuracy of apple leaf disease classification and fruit segmentation in a complex orchard environment,provides technical support for the development of smart agriculture,and advances the development of modern information technology in agricultural problems.