Research On Wheat Detection Based Algorithm On Deep Learning

In the face of the challenges posed by the dramatic increase in population and the frequent occurrence of severe weather,food security has become an important support for maintaining people’s well-being and national security.As one of the three major grains in the world,wheat testing plays an important value in achieving growth and growth stage monitoring as well as yield prediction.However,existing wheat detection technologies face challenges such as different growth stages,intensive targeting,high overlap,and weather and light interference.This has led to problems such as high leakage and false detection rates and insufficient model robustness.To this end,this paper will focus on the above problems in wheat detection algorithm research,with the overall goal of building a wheat detection model with high detection accuracy and robustness,according to the characteristics of the wheat data set,respectively,from the construction of dual color space local-global detection model,the construction of dual-branch wheat detection model based on LSD transformation,and the construction of robust detection model.The specific research contents and main innovation points can be summarized as follows:(1)A dual-color space local-global wheat detection algorithm is proposed for the problem of wheat color diversification caused by different regions and growth stages.The method is based on local feature extraction in RGB and HSV space,and then utilizes Transformer’s long-range modeling capability to achieve local-global multi-view deep semantic information mining,which can obtain richer wheat features than single color space,and thus can effectively improve the detection performance of wheat in different regions and growth stages.(2)A two-branch wheat detection algorithm based on the LSD transform is proposed for the characteristics of wheat ears with rich structural information.The method first extracts the wheat structure information based on LSD transform,then extracts the features from the original and structure maps by building a two-branch network,and finally designs a TFPN structure for multi-scale feature depth fusion,which can significantly improve the detection performance of wheat in dense and obscured situations through the guidance of structure information.(3)To improve the anti-interference capability of the detection algorithm,a robust wheat detection method based on adversarial training is proposed.The method first conducts regular training of the wheat detection method after training In addition,this paper also discusses the adversarial training-based saliency target detection method and the experimental results based on five typical data sets show that the method can also effectively improve the robustness of the saliency object detection algorithm.