Green Fruit Detection And Segmentation Based On Foveabox

The application of artificial intelligence frontier theory injects new connotations into the development of intelligent agriculture,realizing intelligent and scientific management of all aspects of agricultural production,thus significantly improving the efficiency and quality of agricultural production.Exploring more efficient intelligent orchard management has become a hot research,especially for unstructured orchards,the accurately recognizing green fruits is an important part of orchard refinement management.Taking apples in the growing season as the research object,this thesis focuses on the detection and segmentation of green fruits,and discusses efficient and high-precision models based on optimized Foveabox network,mainly as follows.(1)To address the problem that small-scale target fruits are easily missed,a multi-scale fast detection model Fast-FDM is designed.The lightweight feature extraction network EfficientNetv2-S is used to extract features;BiFPN is adopted to fuse multi-scale features and input them to the Fovea head network for classification and bounding box prediction;and the ATSS is employed to alleviate the positive and negative sample imbalance problem for improving the recall rate of green apples at different scales.The experimental results show that with fewer parameters and FLOPs,the average detection precision of Fast-FDM reaches 62.3%.(2)For the boundary ambiguity problem of green fruit segmentation,the boundary-aware instance segmentation model FBSM is proposed.BFS module is constructed in a dual fusion mode at each stage to achieve sufficient fusion of fine-grained features to guide instance mask prediction;the lost boundary detail information is recovered using the BRM structure for the continuous optimization of the instance mask.The experimental results show that the average segmentation precision of the FBSM model reaches 60.7%,and the detection precision is also improved by 1.3%.(3)To alleviate the problem of higher computational complexity of the two-stage segmentation model,the coordinate attention segmentation model SE-COTR is constructed to enhance the focus on target features by combining the MobilenetV2+COTR structure.The improved JPU is used to integrate multi-scale features to cope with different scales of target fruit;and the dynamic convolution operation is applied to predict the instance mask by combining the output of the head network.Experimental results show that SE-COTR achieves an average segmentation precision of 61.6% and 43.3% for small target fruits with low model complexity in the case of heavily obscured and scale-variant target fruits.In summary,for the highly complex and variable nature of unstructured natural orchard environments,this thesis optimizes the model in terms of precision,efficiency,anti-interference capability and module performance of the target fruit detection and segmentation model by applying cutting-edge deep learning theory,effectively solving the difficulties caused by factors such as branch and leaf shading,similar colour and scale differences.This study provides useful reference for intelligent and scientific management of orchards,and also provides theoretical guidance and technical support for applications such as early fruit yield measurement and green fruit machine harvesting,which has important practical application value.