Research On Recognition Methods Of Pomelo Fruit Hanging On Trees Base On Machine Vision

In order to accomplish the picking task,the picking robot needs to recognize the fruit target in the complex background,so the recognition ability was one of the main technical indicators to judge the performance of the picking robot.In this paper,the identification methods of pomelo fruit,hanging on trees which were studied in natural environment.Compared with the performance of several recognition methods under complex background,the recognition ability and portability were selected as the evaluation index to select the optimal recognition method,and it would be used as the recognition module algorithm of picking robot.The main research contents and conclusions were as follows:Data acquisition and partition.A total of 1300 pomelo images were captured by data acquisition method,three intervals were selected as the shooting distance.Divided into 3 shooting intervals: close-range shooting interval was [0.1-1.0m],medium-range shooting interval was [1.0-2.0m],long-range shooting interval was [2.0-3.0m].The pomelo fruit growing in the direction of light and back light were selected as the shooting objects.The shooting fruits included unshielded fruits,fruits that growing in the shade of branches and leaves,fruits that growing in overlap et al.In this paper,a total of 60 fruit images under 2 growing states and 3 shooting interval were selected as verification sets.1010 pomelo fruit images were selected as training and testing sets for deep learning.Pomelo fruit recognition method based on image segmentation algorithm.Chromatic aberration algorithm and K-means clustering algorithm were selected as image segmentation algorithms,the 2 methods were used to segment the images in verification sets in this paper.Chromatic aberration algorithm has a better segmentation effect with the chromatic aberration component of 1.6R-G-B in verification sets;K-means clustering algorithm has a better clustering effect when K=4 was selected.After segmentation,the pomelo fruit region map was obtained by the 2 segmentation methods.Combined with the Ostu algorithm,the area of the connected region was calculated,then the small interference area was removed.At last the smallest outer rectangular box was delineated for the connected region,which was considered as the target region.By statistics the verification sets,recognition results were shown that the F1 measure of chromatic aberration algorithm was 50.07%;F1 measure of K-means clustering algorithm was 62.23%.Pomelo fruit recognition based on deep learning algorithm.Faster RCNN algorithm,which was the target detection algorithm based on candidate region method;YOLOv3 algorithm,which was the detection algorithm based on regression method.In this paper,1010 images in the training and verification set were calibrated as training labels.Faster RCNN and YOLOv3 algorithms were used to train the label images in order to obtain the deep learning algorithm model.The detection and recognition models based on the 2 methods were obtained through training.The loss function values were recorded by the increased iterations.Statistics showed that the loss values were stable in a lower range under the 2 methods.The 2 models were used to verify the accuracy of the verification set.The results were shown that the F1 measure of Faster RCNN algorithm model was 81.74%,the detection time of a single image was 1.09s;the F1 measure of YOLOv3 algorithm model was 88.14%,and the detection time of a single image was 0.60 s.The research results show that machine vision combined with deep learning algorithm could acquire pomelo fruit information in natural background,which provides a technical basis for the subsequent research and development of automatic picking equipment.