The Research On Recognition Of Dendrobe’s Growing Stage Based On Deep Learning

China is a country with a large population and has a large-scale agricultural industry,so the agricultural productivity is vital to the stable development of our country,and obviously intelligentization is an important breakthrough of improving the current agricultural productivity of our country.In recent years,with the rise of smart-agriculture,intelligent methods such as intelligent early-warming of plant diseases based on image and judgement of plants’ s maturity based on image,are beginning to take shape,thus,conclusions can be drawn that computer vision technology plays an important part in the process of the intelligentization of agricultural production.In this thesis,a research is made on the use of Object Detection algorithms on the background of greenhouse-planting dendrobes,and the main contents of this thesis are listed as below:Firstly,dendrobe images are collected and a dendrobe data set is made.Considering both at the greenhouse environment and the features of dendrobe itself,data augmentation methods are designed to prevent the interference from the greenhouse background and dendrobe itself.Secondly,mathematical statistics and analysis are made on the annotations of dendrobe data set,which depicts the morphological characteristic of dendobes in different growing stages with the machine learning K-Means method,and the anchors of network model will be redesign with this statistics and analysis result as guidance.Thirdly,the Object Detection algorithm based on Deep Learning is introduced.YOLOv3,which is a typical one-stage Object Detection algorithm,and Cascade R-CNN,which is a typical tow-stage Object Detection algorithm,are compared over the complicated dendrobe growing stage detection task’s applicability issues,and after some experiments are done,Cascade R-CNN is decided as a better solution for the detection task of dendrobe’s growing stage.Finally,with the disadvantages of Cascade R-CNN analyzed and the dendrobe detection task background considered,some corresponding improvements from two aspects are thus made,including: the improvement of backbone based on Res Ne Xt,the improvement of neck based on PA-FPN,the improvement of anchor based on the K-Means results and the improvement of post-processing based on soft-NMS,and eventually based on all these improvements above,an Object Detection algorithm based on multi-method improvements is put forward,which is utilized to do the detection task of dendrobe’s growing stage.In this thesis,an algorithm suitable for the detection task of dendrobe’s growing stage is designed on the basis of feature analysis of the dendrobe itself,after that some experiments are done to prove the precision of the algorithm,and finally the algorithm is implemented on the visual interface.