| Orange(Citrus sinensis)is an important fruit and cash crop.The detection of condition information for immature orange fruits is the basis for formulating orchard management strategies,and is of great significance for the modern and intelligent orchard.In response to the demand for automated and large-scale detection of immature orange fruits conditions,in this work,a detection method based on deep learning for immature orange fruits and pest infested oranges was proposed.The main research content and achievements are as follows:(1)The background removal methods for immature orange fruits.To address the problem of difficulty in traditional recognition methods due to the similarity between the color of immature orange fruits and the background color,a semantic segmentation model Seg Net was proposed to remove the background of the image to improve recognition accuracy.A dataset of immature orange fruits was established,and then the data was processed using three semantic segmentation models: PSPNet,HRNet and Seg Net.The results showed that the Seg Net model had higher segmentation accuracy than PSPNet model and HRNet models,with 81.72%,82.06%,81.89%,81.17% and 81.74% as the accuracy,recall,F1 score,MPA value and MIo U value,respectively.The Seg Net model was improved based on the learning rate optimization strategy,and the various indicators of the improved model were 83.35%,84.79%,84.06%,82.34% and 82.09%,respectively.Compared with the original model,each indicator had increased by 1.63%,1.73%,2.17%,1.17% and 0.35%,respectively.The established model achieved high-precision of background removal for immature orange fruits.(2)Establishment of a target recognition model for immature orange fruits.A dataset was created,which includes 1500 images with back-ground removed through semantic segmentation.Divide the training and testing sets in a 3:1 ratio.Manually annotated the images and expanded the dataset using data augmentation methods to complete the construction of the immature orange fruit dataset.The Faster RCNN network,SSD network,and single stage detection model YOLOv5 s were used to train the immature orange fruit dataset.The results showed that the YOLOv5 s model had better comprehensive performance,with evaluation index accuracy,recall rate,m AP value,and average recognition time of 94.24%,96.98%,98.21% and 0.067 seconds,respectively.The YOLOv5 s model was improved and optimized based on the hollow space convolutional pooling pyramid structure,and the results showed that the various indicators of the improved model were 96.82%,96.20%,98.93% and 0.071 s,respectively.The recognition rate of images processed by semantic segmentation met the target requirements.(3)Detection of damage to immature orange fruits based on deep learning.1200 datasets consisting of orange anthracnose,orange canker,orange blackspot,and fresh fruits were constructed in a ratio of 1:1:1:1 based on the common damage of immature orange fruits.The dataset was trained using the unimproved YOLOv5 s model and the improved YOLOv5 s model based on attention mechanism.The results showed that the average accuracy,average recall,and average m AP values of the improved model were improved by 2.5%,1.6% and 1.1% compared to the YOLOv5 s model,respectively.(4)Design of the interface for detecting diseases and pests in immature orange fruits based on the Py Qt5 platform.In order to display the information of diseases and pests in immature orange fruits visually,a detection interface for oranges affected by pests and diseases based on the Py Qt5 platform was designed,which can display image size,fruit pest or disease categories;and grayscale value information.The research results indicated that the immature orange fruit recognition method and detection interface proposed in this work can provide information on the quantity and quality of immature orange fruits for orchard management. |
