Research And Implementation Of Cotton Boll Semantic Segmentation And Cotton Intelligent Analysis Platform

Cotton is one of the major economic crops in China and worldwide.Although planting techniques and methods have been widely mechanized,there are still phenomena such as relying heavily on experience to plant,excessive use of fertilizers and pesticides,poor quality of lint,and high economic and labor costs.Therefore,how to fully utilize information technology to achieve intelligent cotton planting has become an urgent problem to be solved.Thanks to the rapid development of computer vision,various AI algorithms based on computer vision have been developed and applied in various stages of cotton growth.cotton intelligent analysis algorithms designed for possible problems during cotton growth effectively reduce the burden on cotton agricultural workers.During various stages of cotton growth,the cotton boll stage is the most vigorous period of cotton growth and development,and it is also the period when cotton requires the most water and fertilizer.Improving the number of bolls and the weight of cotton bolls during the cotton boll stage is a key period to determine the yield and quality of cotton.However,difficulties arise in identifying cotton bolls due to the mutual occlusion between cotton bolls and cotton leaves and the similar colors of cotton bolls and cotton leaves.This thesis proposes a cotton boll semantic segmentation model based on deep learning,and proposes Improved UPerNet for the deficiencies in the UPerNet network.Improved UPerNet uses the ConvNeXt convolutional neural network as the backbone network,and replaces the Pyramid Pooling Module(PPM)in UPerNet with the Atrous Spatial Pyramid Pooling(ASPP)module.Based on the efficient channel attention(ECA)module,this paper improves the filling method and adds the global max pooling branch,proposes Improved ECA(IECA)module,and uses the IECA module to adjust the output of the Improved UPerNet backbone network.The performance of Improved UPerNet in the cotton boll semantic segmentation task is not only better than the classic semantic segmentation model,but also better than the Transformer-based semantic segmentation model.With the emergence of various AI algorithm frameworks,the cost of developing and training AI algorithms is getting lower and lower.However,there is little research on how to deploy AI algorithms on a large scale.For cotton intelligent analysis algorithms,the developers and users of the algorithms belong to completely different fields of work.Trained cotton intelligent analysis algorithms often cannot achieve their desired effects due to the difficulty of deployment and maintenance.Therefore,this thesis designs and develops a cotton intelligent analysis platform,which uses message middleware to decouple the function development module and the AI algorithm module.The function development module adopts the microservice method to develop the interface of each cotton intelligent analysis algorithm.Each cotton intelligent analysis algorithm completes algorithm deployment by listening to the message middleware.The system provides a visualized page for cotton agricultural workers to call cotton intelligent analysis algorithms through front-end and back-end separation development.This thesis also proposes an algorithm management solution based on a Linux server.The system completes the operation and management of the algorithm executable program through the SSH session client.After maintaining the relevant information of the deployed algorithm,the administrator can manage the algorithm on the visual page.