| As an important economic crop,cassava has a planting area of 285,700 hm~2in China,of which Guangxi cassava planting area and output both account for more than 70%of our country.At present,cassava harvesting in Guangxi is still mainly manual,the mechanized harvest is low,and the production cost of cassava is high,which seriously restricts the development of cassava industry in guangxi.For this reason,Guangxi University has developed a new type of bionic digging cassava harvester in recent years.This machine has the characteristics of low cassava root harvesting power consumption and high harvest efficiency.However,due to the lack of accurate positioning of cassava stems,the clamping device cannot accurately grip the stem,resulting in harvest loss.Therefore,it is of great significance to carry out research on the accurate positioning of cassava stems to improve the net yield and intelligent level of the harvester.This articl takes cassava stems as the target,constructs a cassava stems image data set,and uses deep learning-based target detection algorithms for detection research.The main work is as follows:(1)In this articl,on the indoor test platform,simulating the growing environment of cassava in the field,1000 images of cassava stems were collected.Through geometric rotation,adding noise and elastic change data amplification methods,the sample set image is amplified to 11,000.Finally,the sample set was manually marked and divided into 10,000 images of the training set and 1,000 images of the test set.(2)In order to achieve fast and accurate positioning of cassava stems,the YOLO(You Only Look Once)algorithm based on deep learning,while using the method of transfer learning,and removing the Dropout layer,were tested.When the IOU threshold is 0.5 in the test phase,the model accuracy rate is99.5%,which meets the accuracy requirement,but the average detection time of each image is about 0.047s,which cannot meet the real-time detection requirements.(3)In order to improve the detection rate,the Fast YOLO algorithm is used for the detection test.When the IOU threshold is 0.1 during the testing phase,the model accuracy rate is up to 99.1%,which satisfies the accuracy rate requirement,but the average detection time per image is about 0.028s,which basically only meets the real-time detection requirements.The main reason is that Fast YOLO network does not use 1×1 convolution and retains the full connection structure,which leads to the insignificant improvement of detection rate.(4)In order to further improve the detection rate,the YOLO network structure was improved by using the global average pooling to replace the fully connected layer,and the network depth and width were appropriately adjusted,and a new network was designed.The test results show that the size of the new network model is reduced by about half,the average detection time of each image is about 0.015s,the detection rate is significantly improved,and it has fully met the real-time detection requirements.When the IOU threshold is 0.1during the test phase,the model accuracy rate is up to 99%,which satisfies the accuracy rate requirement.Finally,the positioning errors of each network model are analyzed,and all meet the harvesting needs.This research provides a new idea and method for real-time and accurate detection of cassava stems in the field,and provides technical support for bionic cassava harvester. |
