Research On Branch Detection And Parameter Extraction Of Passion Fruit Trees In Natural Environments

Fruit trees are an important part of agricultural production and the accurate acquisition of information on fruit tree branches is of great importance in applications such as monitoring the growth status of fruit trees,automatic pruning and harvesting.Therefore,the detection of fruit tree branches has become a research focus in the field of agricultural informatization in recent years.In early research,the detection of fruit tree branches was mainly based on traditional image processing techniques,such as color space transformation,edge detection,threshold and cluster segmentation and morphological processing.Although these methods are simple and easy to implement,they are susceptible to interference from factors such as lighting and shadows in practical applications,making detection results difficult to achieve satisfactorily.Another method is based on multimodal data,using laser scanning,stereo vision,multispectral and other multimodal data for fruit tree branch detection and reconstruction.This method can describe the morphology and structure of fruit tree branches more accurately,but it requires expensive equipment as well as high-quality data.In recent years,deep learning techniques have been widely used in the fields of computer vision and image processing,and some research results have been achieved in the detection and reconstruction of fruit tree branches.Researchers have started to apply deep detection models(e.g.,Deep Lab,Seg Net,Mask RCNN,etc.)to the detection of fruit tree branches.Although these methods can achieve high accuracy and robustness,they are mostly targeted at deciduous tree fruit trees with relatively simple backgrounds.It is also difficult to detect and reconstruct fruit trees with irregular and dense overlapping branches.This paper focuses on the needs of branch pruning and fruit harvesting,and investigates the automatic recognition and extraction of phenotypic parameters for branches on trees with dense overlapping branches in the canopy and branches with similar colors to other backgrounds under natural conditions in orchards,with the following main work:First,a fruit tree branch detection and segmentation method based on improved YOLOv5 algorithm was proposed for passion fruit trees with small,curved branches and similar background colors.The main optimization measures include: adding a shallow detection layer on top of the original three detection layers,and introducing a new bounding box similarity measurement method called NWD,combined with the original CIo U loss function,to improve the algorithm’s ability to detect small branches.Then,the Sim AM attention mechanism is integrated into the feature extraction module of the original algorithm,allowing the algorithm to focus more on target information and enhancing its feature extraction ability,addressing the issue of similar colors between branches and background.Finally,the CARAFE optimization is used for feature upsampling,solving the issue of information loss in image segmentation with high-resolution feature pyramid.Experimental validation was conducted on images of passion fruit trees grown in a curtain-style plantation.Among them,the precision,recall and F1 score of branch detection are 66.2%,56% and60.67,respectively,and the precision,recall and F1 score of branch segmentation are 69.5%,57.7% and 63.05%,respectively.The results showed that the proposed improved algorithm can accurately detect and segment passion fruit tree branches in natural environmental conditions,making it applicable in practical agricultural scenarios.Second,this paper proposes a branch reconstruction method based on a bidirectional sector search strategy.The method can adaptively reconstruct passion fruit tree branches with minimal parameter adjustments,obtaining complete branch information and fitting the reconstructed branches.Reconstruction accuracy,as well as phenotypic parameters such as branch length and width,were calculated and compared with manually measured data through comparison and correlation analysis.Experimental results show that the accuracy of branch reconstruction is 88.8%,the reconstructed m IOU is 83.44%,and the average relative errors of branch length and width are 4.63% and 7.98%.Experiments validated that this branch detection and reconstruction method has a high reconstruction accuracy and robustness for passion fruit tree branches.