Research On The Online Identification System Of Citrus In Natural Environment

With the launch of the "14th Five-Year Plan" and the full implementation of the rural revitalization strategy,the rapid development of smart agriculture is imperative.Citrus is an important agricultural by-product in our country,and it is widely grown in the south.The mechanical operation of citrus production,especially the mechanized and automated citrus picking operation,is an important factor restricting the development of the citrus industry.Computer vision technology for citrus target detection is an indispensable technical means to realize automatic citrus picking.Traditional target detection algorithms rely on artificial design features,which are time-consuming and labor-intensive,with poor robustness and low efficiency.With the development of science and technology,deep learning technology has gradually become the mainstream of target detection.Deep learning can actively learn the characteristics of the target,and it has a huge improvement in accuracy and real-time.However,citrus picking in a natural environment has a series of complex external factors such as uneven light,overlapping fruits,and occlusion of branches and leaves,which greatly increases the difficulty of online citrus detection.Most of the existing target detection algorithms are trained on ideal data sets,and the adaptability of the algorithms is low in complex natural environments.This paper compares and analyzes the two-stage representative algorithm Faster-RCNN and the end-to-end one-stage algorithm SSD and YOLO.Taking into account the actual needs of automated citrus picking,the YOLO algorithm that merges accuracy and real-time is selected for this research Based on the basic algorithm,the YOLOv4 algorithm is improved and optimized for the problems of target overlap,occlusion,and small targets in citrus picking,and an online citrus recognition system is built to verify the effectiveness of the algorithm through multiple sets of citrus images in natural environments.The main work of this paper is as follows:(1)Improve the YOLOv4 algorithm,which mainly optimizes YOLOv4 from three aspects: network structure,anchor box clustering,and target loss function.In view of the missed detection of small targets,152×152 large-scale feature fusion is added;the anchor box clustering algorithm is changed from the K-means algorithm to the K-means++ algorithm,and the k-means++ clustering algorithm with less randomness is used.The anchor box is used for clustering,which effectively reduces the clustering deviation caused by the original algorithm at the initial clustering point;the loss function is improved to increase the error sensitivity of the bounding box loss function.(2)Make a citrus data set,and when collecting images,ensure that the selected image samples have a better balance and diversity.Label Img is used to label the data set,and a rectangular frame is used to circumscribe the outline of the citrus target.In order to obtain a good detection effect,in addition to using a better model,but also to avoid the model from getting into overfitting,data enhancement is a general practice that can improve the robustness of the algorithm without reducing the detection accuracy.The average accuracy of the improved YOLOv4 model is 84.2%,which is an increase of 1.8% compared with the original network.The experimental results show that the algorithm in this paper has a higher improvement in the accuracy of citrus recognition.(3)An experimental platform for the online citrus identification system was built,and the corresponding online identification system software was designed and developed to realize the online identification of citrus in the natural environment.The citrus recognition test was carried out in the natural environment to verify the feasibility and superiority of the improved algorithm.