Research And Improvement Of Lemon Detection Algorithm Based On YOLOv3

Efficient fruit object detection algorithm is a key technology in fruit picking robots.At present,due to the influence of natural conditions such as different light,the shade of branches and leaves on the fruits,the overlap between multiple fruits,and the color similarity between the fruits and the stems and leaves,the detection of fruits is still a challenging task.In this thesis,the lemon fruit in natural environment is the main research object.In order to improve the accuracy and real-time of fruit object detection,two improved lemon detection algorithms based on YOLOv3 are proposed.The main research work of this thesis is as follows:(1)Construction of the lemon image dataset on trees.First,the lemon fruit images are collected in the lemon plantation from different distances and directions under different weather conditions;secondly,some common image transformation methods are used to preprocess the lemon images to increase the diversity of the images;finally,the lemon images are annotated according to the format of the PASCAL VOC dataset,and the labeled dataset is randomly divided into a training set and a test set,which are respectively used for training and testing of the object detection models.(2)Improved algorithm for lemon high-precision detection based on YOLOv3.On the premise of not sacrificing the real-time performance of the algorithm,in order to increase the detection accuracy,YOLOv3 algorithm is studied and analyzed,and the Lemon-YOLOv3 algorithm is proposed on this basis.In the backbone network,referring to the ideas of Squeeze-and-Excitation Network(SENet)and Res Ne Xt,the SE＿Res GNet34 network is designed to replace the Dark Net53 network in YOLOv3.By enhancing the feature reuse of the model,the model volume and calculation amount are reduced,and the model has higher detection accuracy and speed.In the multi-scale detection module,some convolutional layers are changed to the Res2 Net modules to strengthen the extraction of multi-scale features and further improve the detection accuracy.The experimental results show that the average accuracy of the proposed Lemon-YOLOv3 algorithm on the lemon test set is 96.61%,the detection speed is 99 FPS,and the model volume is 100 MB.Compared with YOLOv3 algorithm,the accuracy and speed are increased by 4.04% and 29 FPS respectively,and the model volume is reduced by 135 MB.(3)Improved algorithm for lemon fast detection based on YOLOv3.In order to make the lemon detection algorithm suitable for mobile or embedded devices,the Fast-YOLOv3 algorithm is proposed based on the lightweight network,which reduces the model volume and speeds up the detection speed as much as possible without sacrificing the detection accuracy.The lightweight network termed Mobile Net V2 with fast object recognition speed and high accuracy is used to replace the backbone network of Dark Net53 in YOLOv3 algorithm,to complete the feature extraction of lemon images,enhance the effective transmission of multi-layer feature information,greatly reduce the model volume,and accelerate the object detection.The multiple convolutional layers in the multi-scale detection module are changed to the M-Res2 Net modules to strengthen the multi-scale feature extraction capabilities of the model and improve detection accuracy.The Balanced L1 loss and the Io U loss are used to improve the YOLOv3 loss function,realizing more balanced training between classification and location,and enhancing the precision of object location.The experimental results demonstrate that the average accuracy of the proposed Fast-YOLOv3 algorithm on the lemon test set is 94.39%,the detection speed reaches 159 FPS,and the model volume is 44 MB.Compared with YOLOv3 algorithm,the accuracy and speed are increased by 1.82% and 89 FPS respectively,and the model volume is reduced by 191 MB.(4)In order to verify the effectiveness and applicability of the improved algorithms in this thesis,different object detection algorithms are used to conduct comparative experiments on the homemade lemon image dataset,the homemade mango image dataset and the public field grape image dataset.The results show that,compared with other state-of-the-art object detection algorithms such as Faster R-CNN,Cascade R-CNN,Single Shot Detector(SSD)and YOLOv4,the proposed Lemon-YOLOv3 and Fast-YOLOv3 algorithms better balance the accuracy,speed and model volume of detecting fruit objects.The two proposed algorithms can meet the task of high-efficiency detection of fruits.In summary,the Lemon-YOLOv3 and Fast-YOLOv3 algorithms proposed in this thesis have strong detection performance,which can recognize and locate lemon fruits in natural environment faster and more accurately,and provide a reference for realizing the machine picking of lemons and other fruits.