Research On Real-time Detection Algorithm Of Passion Fruit Maturity Via Deep Learning

With the continuous improvement of agricultural intelligence in China,automatic picking technology has become an important part of agricultural picking robots.In recent years,real-time detection of picking objects in automatic picking technology tasks of picking robots has become a research difficulty.For improving the precise classification and positioning of picking objects by the picking robots,combined with the actual problems passion fruit maturity detection faces in the natural environment,deep learning and technology related to the image processing are used to do real-time detection research of passion fruit maturity,and the detection system of passion fruit maturity is designed.The main research contents are as follows:This paper introduces the research status of fruits target detection,analyzes the principle of convolutional neural network,lists two common target detection algorithms based on candidate area and regression.In addition,passion fruit dataset are collected,pre-processed and produced.In order to solve the limitations of traditional fruit target detection,a real-time detection algorithm of passion fruit maturity based on RetinaNet is proposed.This research analyzes the basic principle of RetinaNet algorithm,which extracts feature based on Res Net,uses Feature Pyramid Networks(FPN)for information fusion of multi-scale feature,does target objects classification and location through classification subnetwork and regression subnetwork.For further improvement of detection accuracy,the RetinaNet is optimized.The Convolutional Block Attention Module(CBAM)is added in the classification subnetwork to enhance the extraction of characteristic information to improve the classification accuracy of the network.Optimization algorithm Adaptive Moment Estimation(Adam)is also used to optimize the parameters of the network model and speed up the training of the model.Through loss curve,average precision curve and PR curve,combined with comparison of various algorithm detection results,the excellent performance of proposed algorithm is verified.In order to further solve the problem of detection accuracy and real-time performance of passion fruit maturity in natural environment,a real-time detection algorithm of passion fruit maturity based on YOLOv3 is proposed.The basic principle of YOLOv3 algorithm is explained through DarkNet53 backbone network,K-means clustering to label anchor box values and multi-scale fusion prediction.Aiming at passion fruit small target detection and reducing the generation of missed and false detections,YOLOv3 is optimized through eliminating large object forecast scale of YOLOv3 and adding Densenet network after medium object forecast scale to enhance the network characteristics to improve the detection precision of the model.At the same time,the Soft-NMS(Non-Maximum Suppression)method is used to screening predict box.The reslut shows the average accuracy of the improved YOLOv3 algorithm reaches 0.906,0.0828 higher than the original algorithm,and the detection speed is 32 FPS,which meets the real-time detection of passion fruit maturity in natural environments.The passion fruit maturity detection system is designed after analyzing the practical requirements for passion fruit maturity detection.This system has embedded the trained RetinaNet-based and YOLOv3-based algorithm model for classification and localization of passion fruit maturity pictures respectively.Meanwhile,the function of data pre-processing is expanded,the interface and each functional module of the system are elaborated.Experimental results show that proposed system has good stability,simple operation,convenient use and other practical application values.