Accurate Detection Method And Application Of Banana Tree For High Resolution UVA RGB Images

Crop detection is important for yield estimation,weed control,phenotype or disease management,all of which are relevant to the productivity of orchards.With the development and popularity of UAVs,it has become possible to use UAVs to collect high-resolution RGB images combined with deep learning algorithms to achieve accurate detection of banana trees.The following is the main research of this article:(1)Constructing a dense target detection model for banana tree images.A total of 552 dense banana tree remote sensing RGB images were collected,and part of the dataset was used for model training and part was used to synthesize highresolution banana tree panoramic images.After several different rounds of model training,a model with an average accuracy(m AP)of 87.29% was obtained,which has good application prospects.(2)Designing a high-resolution banana tree image cutting reduction algorithm.To address the shortage of YOLOv4 that cannot detect high-resolution images,the high-resolution images are cut into subgraphs by a certain cut size and cut overlap rate with variable sliding windows,and then the subgraphs are detected,and finally the detection results are restored to the original images.The detection time of the algorithm is inversely proportional to the cut size and positively proportional to the cut overlap rate,and the m AP reaches at least 55% and up to 78.13%.(3)Designing a de-duplication algorithm for high-resolution banana tree dense target detection.To solve the banana tree repetition detection problem,MIo P values are added as constraints based on the traditional NMS de-duplication algorithm.The improved MIo P NMS algorithm reduces the counting error to within 7% overall,and the m AP rises by at least 7.33% and at most 28.93%.The MIo P NMS algorithm was evaluated for different training rounds of the model,and the counting error was reduced by anywhere from 8.7% to 23.7% compared to several commonly used NMS.The algorithm has good robustness.(4)Development of an accurate statistical system for banana tree population.In order to realize the engineering implementation of the above research results,a high-resolution banana tree number accurate counting system is developed by using Py Qt.The accuracy of this system is reflected in: firstly,it can accurately divide the banana tree detection and counting plots,and secondly,it can accurately detect the banana tree location and quantity information.And the detection and counting efficiency of this system is 78 times higher than that of manual counting.It has good application prospects.