Research On Smart Vision Algorithm Of Table Grape Inflorescence And Berry Shaping

China is currently the world’s second-largest grape producer and the largest producer of table grapes.With increasing purchasing power and adjustments in dietary structure,there has been a rapid growth in consumer demand for high-end grapes,particularly for premium table grapes that are visually appealing,brightly colored,and have uniform-sized berries.In the production process of premium table grapes,the shaping of grape inflorescences and berries is crucial.Currently,the shaping of table grapes relies mainly on manual labor,which is inefficient,labor-intensive,and unable to meet the demands of large-scale vineyards.There is an urgent need to design a grape shaping robot with a simple structure and high stability.Achieving automation and intelligence in the grape shaping robot requires a key component—smart computer vision algorithms for grape shaping.However,during the thinning period,table grape inflorescences have small targets with complex poses,and during the thinning period,grape berries are densely packed and mutually obstructing,severely affecting the detection and positioning of grape inflorescences,stems,and berries.This poses a significant challenge to the smart computer vision algorithm for grape shaping.The precise and rapid localization of thinning points on table grape inflorescences and the identification of grape inflorescences and berries to be removed are the foundation of the smart computer vision algorithm for grape shaping and are crucial for the grape shaping robot.Therefore,to meet the agricultural and mechanization requirements of grape shaping,corresponding smart computer vision algorithms for table grape shaping are proposed.The main research content of this paper is as follows:(1)A model for segmentation of grape inflorescences and localization of thinning points was proposed based on the Mask R-CNN architecture to achieve precise positioning of the thinning points for table grape inflorescence,The model employed the ResNeXt network as the backbone and incorporated path enhancement to address the challenge of detecting small targets such as table grape inflorescences and stems.Considering the complexity and uncertainty of the growth posture of grape inflorescences and stems,a collective logic algorithm was introduced for the precise localization of thinning points.Experimental results demonstrate that,under different scenarios,the average positioning accuracy of thinning points for table grapes was 83.3%,with an average localization time of 0.325 seconds.The maximum positioning errors in the x and y directions,as well as the total positioning error,were 10,12,and 16 pixels,respectively.These findings indicate that the model meets the precision and speed requirements for the mechanical positioning of thinning points in table grapes,providing theoretical support for the automation of thinning in table grape production.(2)A lightweight model for grape inflorescence detection and thinning point localization was proposed based on the YOLOV7-TP architecture to further reduce the model’s parameter count and achieve rapid positioning of thinning points for table grapes.Different YOLOV7TP models were constructed with varying numbers of thinning points for detection,and the model with the optimal number of thinning points was selected.The impact of hyperparameters on the detection and localization accuracy of the YOLOV7-TP model was compared.Channel pruning was employed to obtain a lightweight model for grape inflorescence detection and thinning point localization,and various hyperparameter combinations were compared to determine the optimal set.Experimental results indicate that the detection accuracy of the model for table grape inflorescences,with a mean average precision(mAP)at IoU 0.5,reached 91.5%.The model’s parameter count was 2.3 million,GFLOPS were 8.7,and the detection speed was 29.4 frames per second(FPS).(3)An algorithm for thinning decisions in table grapes was proposed based on YOLOV5s and K-means to meet the agronomic and agricultural machinery requirements for thinning table grapes and automatically make decisions regarding the removal of individual small spikelets within a grape inflorescence,The YOLOV5s backbone network was enhanced with a Pyramid Split Attention module(PSA)to improve the model’s feature extraction capabilities.The neck network utilized a Bi-directional Feature Pyramid Network(BiFPN)with weighted feature fusion to enhance the detection accuracy of small spikelets.The CIoU(Complete Intersection over Union)loss function was employed to optimize the regression accuracy of bounding boxes.The K-means algorithm was used to cluster the bounding box centers of grape small spikelets,identifying inflorescence tails and determining that 2/3 of the small spikelets should be removed.The hyperparameters of the thinning decision algorithm for table grapes were analyzed.Experimental results show that,with an Intersection over Union(IoU)threshold of 0.15,confidence threshold of 0.6,and 20 small spikelets,the algorithm achieved a maximum accuracy of 78%.(4)It is necessary to further thin the fruit after thinning the blossoms for premium table grapes in accordance with the agronomic requirements of table grape production.To automatically make decisions regarding the removal of grape berries within a single grape cluster,a decision-making algorithm for thinning table grapes was proposed based on ASSwinT and DeepLabV3+.Addressing the issue of dense occlusion of small grape berries during the thinning period,an AS-SwinT model was constructed.Combining this with a linear regression model,a model for instance segmentation and counting of grape berries based on AS-SwinT was proposed to determine the number of berries to be removed.DeepLabV3+ was employed to segment and fit grape stems,obtaining the optimal segmentation model.By considering the number of berries to be removed,the phenotype of the grape cluster,and the phenotype of individual grape berries,the algorithm achieved automatic thinning decisions for table grapes.Experimental results demonstrated an average precision of 94.5%,an average recall of 78.8%,and an average F1 score of 85.8%for this method.The approach exhibited higher accuracy and met the requirements for thinning table grapes effectively.