Detection And Segmentation Of The Same Color Fruits In Complex Orchard Environments

The application of artificial intelligence technology has promoted the innovation of intelligent equipment in orchards,improved the production efficiency and quality of orchards,and realized the intelligent and scientific management of orchards.Fruit monitoring and yield measurement is an important part of orchard intelligent management,relying on the accurate detection and segmentation of target fruits.However,the detection and segmentation accuracy of green fruits is low under the complex and changing orchard environment,which is affected by light changes,overlapping shading and similar background color.In this study,two same color datasets of green apples and green persimmons are produced,and the accurate detection and segmentation models of target fruits of the same color are discussed as follows.(1)To address the problem of low detection accuracy due to uneven scale of target fruits,a Dense to Detection(D2D)model based on anchor is proposed.The lightweight Mobile Net v2 is used as the backbone network to reduce the number of parameters and computational complexity;the dense detection method is used in the regression branch to calculate the offset to precisely locate the target fruit;the classification branch uses adaptive weight pooling for different sub-regions to generate classification labels and confidence levels.The experimental results show that the average accuracy of the D2 D algorithm is 73.4% and 62.8% for the persimmon and apple datasets,respectively.(2)For the anchor frame dependence problem of the detection model,a one-stage anchor-free frame detection model based on FCOS optimization is constructed,and the bottom-up path enhancement module is connected after the Feature Pyramid Network(FPN)structure to integrate different scales of features;the Optimal Transport Assignment(OTA)algorithm is used to assign positive and negative samples of anchor points on feature maps of different scales;a 1×1 convolutional layer is added to the centerness branch and regression branch to output the prediction results of the image and accelerate the model convergence speed.The experimental results show that the average accuracy of the new algorithm is improved by2.7%.(3)To address the problem of computationally intensive and low accuracy for segmenting same colored fruits,we propose an optimized Mask R-CNN instance segmentation model,using Mobile Net v3 as the backbone network to reduce the complexity of the model;the rough mask boundary block is optimized using the Boundary Patch Refinement(BPR)post-processing module to smooth the segmentation contour boundary.Experimental results show that using the BPR module can improve the segmentation accuracy of the target fruit by 2.6%,and the overall optimized average accuracy is improved by 3.1%.In summary,the deep learning-based object detection and instance segmentation models are proposed in this study to solve the problem that the visual system is severely disturbed in complex orchard environments,realizing the efficient detection and segmentation of homochromatic target fruits.The above study enhances the anti-interference ability and recognition accuracy of the model,realizes efficient orchard monitoring and yield measurement,and provides theoretical reference for other fruit detection and segmentation studies,and promotes the development of intelligent equipment in orchards.