Green Fruit Detection And Segmentation In Orchard Environment Based On Improved FCOS Algorithm

Currently,China’s orchard production is characterized by large fruit production,labour shortage and low level of automated operations.With the support of the national policy of vigorously advocating the priority development of agricultural machinery intelligence,accelerating the research and development of intelligent operation equipment for orchards has become an urgent task.The vision system is the core part of the orchard intelligent equipment.Under the complex orchard environment,affected by interfering factors such as light intensity,branch and leaf shading,and the similarity of fruit and background colour,the performance of the vision system is thus affected,resulting in low recognition accuracy of the target fruit.Taking green apples and persimmons in a complex orchard environment as the research object,this study explores in detail the efficient detection and segmentation of green fruits based on the optimized FCOS network,mainly as follows.(1)To address the problem of feature information loss due to scale inconsistency,the efficient detection model FRS-Det is proposed.Based on the FCOS network,a switchable cavity convolution is used to optimize the residual network and expand the perceptual field;a cyclic feature pyramid is constructed by embedding feedback connections and cavity pooling pyramids to fuse features at different scales;and a feature refinement module is added to enhance feature fusion.The experimental results show that the detection accuracy of FRS-Det algorithm on green apples and persimmons achieves 85.1% and 79.4% respectively.(2)Aiming at the unbalanced accuracy and efficiency of fruit segmentation,the efficient segmentation model Polar-Net is designed.Based on the Polar Mask segmentation model obtained by optimizing the FCOS algorithm,Dense Net is used as the backbone network to improve the feature reuse rate and reduce the number of model parameters;the cross-attention module is embedded to construct a cross-feature pyramid structure,focusing on the phenotypic features of the local region of the target fruit.The experimental results show that the segmentation accuracy of the Polar-Net algorithm reaches 87.1% and 81.1% for green apples and persimmons,respectively.(3)For the difficulty of feature representation in the obscured fruit regions,a DLNet dual-attention mechanism segmentation algorithm is constructed.The FCOS algorithm is optimized to design a non-local feature pyramid structure to capture global contextual information and improve feature representation;a two-layer graph attention network is constructed to establish the interaction between occluding and occluded regions and enter different layers for mask prediction respectively.The experimental results show that the segmentation accuracy of the DLNet algorithm achieves 79.1% and 81.2% for obscured green apples and persimmons respectively.In summary,for the detection and segmentation of green target fruits in complex orchard environment,this study combines the advantages of FCOS network to explore the anti-interference ability and generalization ability of the model,which effectively alleviates the difficulties caused by factors such as light changes,similar background colors and overlapping occlusions.It provides theoretical support and technical support for the development of intelligent equipment for orchards,and also provides theoretical reference for the recognition of other fruits.