Efficient Segmentation Of Green Fruits Based On Optimized Anchor-free Method

The production of China’s orchards is facing difficulties such as low efficiency of human operations and shortage of agricultural labor resources,and it is urgent to vigorously improve the intelligence level of orchard automation operation equipment to promote the sustainable development of the orchard industry.Vision system is the core part of the intelligent equipment,and the accurate segmentation method of fruit determines the performance of the vision system.Influenced by the same color and shading of branch background,fruit overlap and uneven scale,exploring the accurate segmentation of green fruits in the complex orchard environment has become a research hotspot and aporia for orchard intelligent equipment.This study takes green apples and persimmons as the research object,and focuses on the efficient and accurate segmentation of green fruits,explores in detail the optimization from anchor-based frame to anchor-free frame methods,with the main research contents as follows.(1)To address the problem of low efficiency of green fruit segmentation,a YOLOF-Snake efficient segmentation model is designed,using only one layer of feature maps based on YOLOF to reduce model complexity;the embedded snake structure is iteratively adjusted for fruit contours to improve segmentation accuracy.The experimental results show that the segmentation accuracy of the new model on apple and persimmon datasets is 79.5% and 76.6%,respectively,and the average segmentation time is 0.42 s and 0.40 s respectively.(2)For the overlapping fruit segmentation boundary blurring problem,the FCOS optimized segmentation model is constructed,and the boundary attention module is introduced in the head network to integrate the rough edge position and classification prediction to accurately detect the fruit boundary;the mask branching and edge segmentation module features are incorporated to model the pairwise affinity of all pixels of the feature map using non-local affinity resolution;prediction results output by fruit shape and appearance common feature map.It is demonstrated that the method achieves segmentation accuracy of 80.5% and 77.9% on apple and persimmon datasets,respectively.(3)To address the problem of feature loss in segmentation of the obscured fruit region,the FCOS network structure is further optimized by introducing the Residual Feature Augmentation(RFA)module,which constructs a residual feature pyramid to reduce feature information loss through spatial context information;the backward segmentation model introduces two overlapping convolutional blocks of attention in the region of interest network layers,in which the top layer detecting the occluded object and the bottom layer reasoning about the partially occluded target fruit.Experiments demonstrate that the segmentation accuracy reaches 81.3%and 78.7% on the apple and persimmon datasets,respectively,and the complexity of the model is further reduced.In summary,for the complex orchard environment of green overlapping fruit and branch and leaf shading fruit segmentation difficulties,this study proposes an optimized segmentation algorithm from anchor-based method to anchor-free method with the help of cutting-edge deep learning theory.And it effectively solves the interference of overlapping fruit,branch and leaf shading,background of the same color and other factors.This study extends the research and development ideas of orchard intelligent equipment,and provides technical support to realize scientific and intelligent management of orchards.Meanwhile,the new model provides theoretical reference for other fruit recognition research.