Intelligent Grading System Of Lentinus Edodes Sticks Based On DeepSORT Tracking Algorithm

Lentinus edodes sticks play a very important role in the process of cultivating high quality lentinus edodes.Good quality lentinus edodes sticks can grow better lentinus edodes.The traditional grading method of lentinus edodes stick was mainly done based on experience which was subjective and fewer traits was extracted.To solve the above problems,an intelligent grading system for lentinus edodes sticks was developed in this study which could extract online the number of lentinus edodes,overlapping rate of lentinus edodes,cover area and color of lentinus edodes sticks,and automatically grading lentinus edodes sticks according to the phenotypic characteristics.The main research contents and results are as follows.(1)Design and development of grading device.The grading device is composed of dark box,light source,rotating device,clamping device,industrial camera,infrared temperature sensor and single chip microcomputer.Dark box provide reliable working environment,and infinite variable light LEDs provided stable light source.The rotating device controlled by single chip microcomputer and it makes the lentinus edodes stick rotate counterclockwise in the dark box.A clamping device above the rotating device can hold the lentinus edodes stick to prevent it from falling off during rotation.Industrial camera captures images of lentinus edodes sticks during rotation and infrared temperature sensor collects the surface temperature and ambient temperature of lentinus edoides sticks in real time.The experimental result shows that the average speed of the rotating device is2.70°/s,and the rotation angle is 157.50°.At this speed,the system can track the lentinus edodes stably.(2)Grading system of lentinus edodes stick based on YOLOv4 and Deep SORT multi-target tracking algorithm.In this study,a lentinus edodes recognition model based on YOLOv4 neural network was established and Deep SORT multi-target tracking algorithm was used to track lentinus edodes on stick during rotation.The anchors of YOLOv4 neural network was reclustered to make it more suitable for the size of lentinus edodes on the stick.The activation function of the YOLOv4 was replaced with Leaky Relu which improved the computation speed in the process of lentinus edodes identification.The feature extractor and tracker of Deep SORT multi-target tracking algorithm were improved to prevent the migration of lentinus edodes id and repeated tracking of the same lentinus edodes.The error of the improved tracking algorithm is 2.85%.At the same time,several phenotypic characteristics of lentinus edodes sticks were extracted during the tracking process.The random forest algorithm was used to grading the lentinus edodes stick based on the extracted phenotypic characteristics and the accuracy of the prediction of sticks in the test set was 99.98%.(3)Design and implementation of software interface.A software interface was designed for the intelligent grading system of lentinus edodes sticks.The software can display the tracking process of fungus sticks in real time,and display the phenotypic characteristic datas and grading results on the software interface.The software can control the rotating device and start the tracking algorithm.The image processing speed is 0.32s/frame,and the average time required to complete a lentinus edodes sticks grading was58.23 s.Py Installer is used to encapsulate the developed algorithm,the packaged software can be used on computers with different performance and no deep learning environment.