Research On Citrus Detection In Natural Environment Based On Deep Learning

The application of deep learning technology in the field of fruit intelligent picking is gradually mature.The key to fruit intelligent picking technology is to achieve real-time stable testing of fruit.Citrus has a wide planting area in my country.In order to improve the degree of intelligence of citrus fruit picking,reduce the workload of fruit farmers,and improve the efficiency of picking,it is important to achieve the rapid and stable testing of citrus fruit targets.This article selects citrus fruits in the natural environment as the research object.For the three key problems of citrus fruits facing citrus fruits in the natural environment,three key problems of citrus fruits,citrus fruits,and surface defects on the surface defects of citrus fruit.Fruit images use different deep learning algorithms to build cittus detection models in natural environments.Improve the speed and detection accuracy of citrus detection in the natural environment,provide a technical basis for the design of the visual detection system of citrus smart picking robots,and provide technical reference for the achievement of fast and intelligent picking of citrus fruits.The research work of this article is as follows:Research on target detection of yellow-green citrus fruits in natural environment.A deep learning model is used to conduct research on yellow-green citrus testing and explore the common model of citrus detection with two color characteristics.Establish the yellow citrus image dataset and the green citrus image dataset in the natural environment.Three deep learning detection models:VGG16-based Faster R-CNN,Resnet-based Faster R-CNN and YOLOv5s were used to compare yellow citrus and green citrus in natural environment.The experimental results show that for yellow citrus,the accuracy,recall,F1 value and AP value of the YOLOv5s model detection method on the citrus test set are 91.9%,99%,0.94 and 97.4%,respectively,and the average detection speed is 32 frames/s.For green citrus,the accuracy,recall,F1 value and AP value of the YOLOv5s model detection method on the citrus test set were 96.5%,98%,0.96 and 97.2%,respectively,and the average detection speed was 27 frames/s.The results show that for citrus with two color characteristics,the YOLOv5s detection model has better versatility and faster detection speed,which is more suitable for citrus detection research.Research on target detection of shielded citrus fruit in natural environment.Based on the target detection of yellow-green citrus fruits,the occlusion images of yellow-green citrus fruits in natural environment were collected,and the detection model was improved by adding attention mechanism.Firstly,four models Faster R-CNN,YOLOv4-Tiny,YOLOv5s and YOLOXs were used for comparison test.The algorithm YOLOv4-Tiny and YOLOv5s with relatively good comprehensive performance under the occlusion data set were selected for optimization,and the algorithm was improved with the attention mechanism.After experimental comparison,in the occlusion data set,compared with the original YOLOv5s model,the average detection accuracy of the improved method with the addition of attention mechanism CBAM increased by 5.14 percentage points,reaching 92.28%,and the detection speed was 36 frames/s.The average detection accuracy of ECA’s improved method increased by 4.99 percentage points to 92.13%,and the detection speed was 44 frames/s.Detection of citrus fruit surface defects on trees in natural environments.Intelligent selective picking of citrus fruit is beneficial to improve picking efficiency.The shape of citrus fruit surface defects is irregular and the lesions are various.In this paper,according to the common defects of citrus fruits,such as canker disease,anthracnose,scab disease,sunburn,mildew and cracked fruit,the defective fruit data set was established,and three semantic segmentation models were selected:Deeplabv3+,Unet and Pspnet models were built using different trunk networks,model training and optimization were carried out,and the lightest Pspnet-Mobilenet model was finally selected as the defect detection model based on various evaluation indexes.The mIoU value of the test was 86.20%.mPA was 92.62%,and the detection speed reached 75 frames/s.