Study On The Recognition Method Of Green Citrus In Non-structural Environment Based On Machine Vision

China's agricultural production has long relied on labor,labor intensity is large but inefficient.The rapid development of computer and automatic control technology has greatly promoted the improvement of automation and intelligence level of agricultural production.Among them,the machine vision technology has high detection accuracy,information rich,non-destructive testing and other characteristics,for the orchard robot positioning picking provides technical support.While,the problem of fruit posit ioning in similar background is the difficulty of machine vision detection technology.Therefore,the research on t HSI problem is helpful to the popularization of machine vision technology in picking application.It has a very important role and economic benefit for agricultural automation and intelligent production.This paper mainly studies the fruit detection technique in similar background,and compare the traditional image segmentation algorithm and the depth learning algorithm with green citrus as the experimental object.The design and parameters of the algorithm are optimized by multiple experiments,and the advantages and disadvantages of the two algorithms are discussed through several sets of comparative analysis.The main contents of the paper are as follows.In the traditional image segmentation algorithm,t HSI paper chooses the threshold segmentation algorithm to carry on the target detection.Threshold segmentation algorithm can be divided into three stages: preprocessing,segmentation and post-processing.In the preprocessing stage,the image color characteristics are analyzed,and it is determined that the contrast between the fruit and the background is large in the Cb component image,and it can be used for the subsequent segmentation operat ion.In the segmentation stage,the paper mainly compares the enhancement effect of several different enhancement algorithms on the contrast of the image.Finally,the paper selects the Retinex algorithm to enhance the image,and uses the Otsu method to se lect the appropriate segmentation threshold.In the post-processing stage,morphology processing was performed on the binary image obtained from the previous step to get the segmentation result.In the depth learning algorithm,tHSI paper chooses Faster RCNN to carry on the target detection.THSI part can be divided into experimental preparation,tuning training and model testing.The experimental preparation is mainly for the configuration of the experimental environment,the production of experimental dat a sets and artificial labeling documents,the selection of the pre-training model and loss function.The main process of tuning training is to experiment with multiple parameter combinations to select the appropriate hyper-parameters,such as learning rate,batch size and momentum,and then use the determined hyper-parameters to train the model.The model testing is to evaluate the detection effect of the trained model on green citrus test set,and the accuracy of the detection is 83.24%.Finally,the paper designs a number of comparative experiments to evaluate the performance of the two algorithms.These comparative experiments mainly study the detection effect of two algorithms on different images.These images include the images with different numbers of fruit,the images with different area scales of fruit,and the mages of different illumination angles.According to the results of the comparative experiments,the conclusion is that the threshold segmentation algorithm is suitable for the locating detection of the picking robot,while the Faster RCNN is suitable for the detection of multiple or small area scale fruit targets.The experimental results of Faster RCNN provides the basis for the feasib ility of the produce estimation.