Evaluation And Classification Of Citrus Fruit Characteristics Based On Instance Segmentation

The citrus plantations in China are mainly distributed in remote mountainous areas,with high labor intensity and low efficiency.The application of intelligent equipment can reduce the difficulty and intensity of every link of orchard production management,especially the equipment based on computer vision,which can replace people to perform related tasks in complex scenes.Orchard intelligent devices based on computer vision need to recognize specific objects from the images captured by the camera before they can perform the next steps.Object detection recognizes objects from images,so it can also be regarded as a classification problem in the field of image processing.The instance segmentation algorithm of object detection can solve the problems of object representation,classification,location and segmentation.An improved algorithm based on the NNC algorithm is proposed.By evaluating the similarity between the training samples and the test samples,the subspace information of the training samples and the test samples is extracted to represent the samples.This algorithm can get better recognition rate.In addition,the algorithm uses fast QR decomposition to solve the least squares problem,and this method is efficient.The test results of ORL and INDIAN face datasets show that the improved algorithm reduces the inter-class distance to about 20%,improves the recognition rate by 24% compared with the NNC algorithm,and increases by about 3% compared with the CNNC algorithm.In order to train the deep learning model of citrus,image samples of citrus were collected and labeled.Mask R-CNN pre-training model was used to fine-tune the collected citrus image,and the detection model of Citrus image was obtained.The results of model test show that the Io U of Mask,which accounts for a large area of image(such as pomelo),can reach more than 85%,while the Io U deviation of small targets(such as leaves)is larger,ranging from45% to 70%.The model detection accuracy of training is lower than that of large-scale open dataset training,so there is still much room for improvement.By optimizing the code,the processing speed can be improved.Using images with high SSIM value,the total processing speed of FPS is about 1～2,which requires a large hardware cost to improve FPS.Therefore,the application of Mask R-CNN in real-time detection is limited.The citrus image segmentation model based on YOLOv3 is superior to Mask R-CNN in speed,but the test results show that the number of detected targets is 30% less than the latter,and the positioning accuracy of the target is at least 10% more than the latter,which is quite different from the YOLOv3 pre-training model.The experimental results show that YOLOv3 needs more sample for training in order to get better detection results.Processing speed test shows that the FPS of this model is about 16～20.It is found that the relationship between FPS and SSIM value is not obvious through static image and video test,but it is related to the number of detected objects,but the dynamic range of FPS of Mask R-CNN is about 18±2,while the dynamic range of FPS of Mask R-CNN can reach 2±1.5 or even lower.The test results show that the processing speed of YOLOv3 is stable on a specific hardware platform.By using data enhancement method,rotation and scaling transformation of the test sample can reduce the distance between the test sample and the training sample,which can effectively improve the number of recognitions.In the experiment,the rotation method was used to increase the detection number by 50%,while the resolution adjustment method was used to increase the detection number by 35%.Data enhancement need to take actual situation into account,otherwise it will affect the detection results.After comparing the test results of YOLOv3 with Mask R-CNN,it is proposed to use Mask R-CNN to build a sample library and YOLOv3 to detect common objects online.Finally,the custom sorting of pomelo is realized based on PCA and k NN/SVM.Intelligent devices need complex data exchange with server.So,this paper designs a cloud platform and remote communication terminal based on NB-Io T.JSON data interaction between cloud server platform and remote devices is realized.It not only compatible with the foundation of traditional devices,but also provides a convenient data channel for remote control of intelligent devices.The innovation lies in:(1)using new representation method to reduce the distance between similar test samples and training samples to 20%,using QR to solve the least squares problem,the algorithm efficiency is improved,the recognition rate is 24% higher than that of NNC and 3% higher than that of CNNC;and(2)combining data enhancement method,the recognition rate and accuracy of the model are improved,among which the detected objects’ s number can be increased by 50% after rotating the image,and by adjusting the resolution,the resolution can be increased 35% detected objects.Based on the results of instance segmentation,the volume calculation,angle calculation and quality evaluation of pomelo are realized.Thirdly,based on instance segmentation and integrating the advantages of several algorithms,a set of classification method based on user-defined classification samples is developed.Using Mask to remove the background,the distance between the test sample and the training sample can be reduced to 20～40%.Finally,the latest Io T technology NB-Io T is applied to mountain orchard to realize JSON data interchange between terminal hardware and Internet platform.A filtering algorithm is proposed to reduce the number of connections and the terminal is compatible with the original peripherals.