Research On Accurate Identification Method Of Fruit Trees Based On High-dimensional Feature Fusion

Fruit trees are the smallest unit of digital orchard management,and accurate identification of the number and spatial distribution of fruit trees is the basic condition for fine farming management of orchards such as growth monitoring and yield prediction.The interference of many factors such as shadows of fruit trees,weeds,other crops and camera imaging conditions causes the image quality collected by UAV to be unstable,which to some extent affects the accuracy of fruit tree detection with a single feature and the applicability of fruit tree detection models.Image segmentation prevents fruit tree targets from being truncated by setting the overlap distance between adjacent images to ensure the integrity of fruit tree target recognition,but this leads to the fruit trees in the overlap region being repeatedly detected and generating multiple detection results.To address the above problems,this thesis proposes a new idea for fruit tree recognition and duplicate result de-duplication statistics,which is divided into the following three aspects.(1)In this thesis,based on the YOLOv5 model and Cross-Modality Fusion Transformer(CFT)model,the YOLOv5-RGBD model is proposed for fruit tree detection by improving the data input interface.For the first time,the joint features of fruit trees in spectral and spatial dimensions are automatically extracted by machine learning in the same deep learning framework,and feature fusion at the raw data level is achieved for fruit tree detection.The experimental results show that the YOLOv5-RGBD model has high accuracy and model recognition performance.The F1 value of citrus tree recognition is 97.3%,which is 3.4% and 1.8% higher than the experiments in which the YOLOv5 model extracts the spectral and spatial features of fruit trees,respectively,and the F1 value of orange tree recognition is 98.4%,which is 7.0% and4.0% higher than the comparison methods,respectively.(2)An innovative design of feature fusion effect validation experiment.The YOLOv5-RGBD model can accurately identify fruit trees even when there is loss of one feature information and the other feature information is intact,which visually confirms the effectiveness of the fusion method.(3)A statistical de-duplication method for the number of fruit trees was proposed.Firstly,the relative coordinates of the detection frame were converted into absolute coordinates on the images of the study area using global projection,secondly,the coordinate results of fruit trees within each image were stitched together,and then duplicate results were screened by simultaneously referring to the combined values of four elements: the predicted value of fruit tree identification results,the confidence of detection frame,the probability of labeled frame width and height,and the planting distribution;finally,the duplicate results were de-duplicated by using the clustering method.The results showed that the final detection accuracy of both citrus tree and orange tree number statistics could reach more than 98%.