| Counting plays an important role in agricultural production and is at the core of a series of tasks such as yield estimation,growth status monitoring,and plant phenotyping.With the advancement of science and technology,many agricultural operations are gradually replaced by machinery.However,field crop counting is still mainly done by manual efforts,which is laborious,costly,inefficient,error-prone and can cause irreversible damage to crops.More importantly,the need of large-scale and high-throughput analyses in modern agriculture makes it impossible to deal with such tasks in a manual manner.Although many automation solutions have been proposed to enhance the production efficiency and liberate labor force in recent years.Most of these methods are designed for plants grown in controlled environment,and cannot be effectively applied to the actual unconstrained field environment.To better address the problem of object counting under unconstrained field conditions,this thesis takes maize tassel as research object and conducts studies based on Computer Vision.First,to address the problem that the current state-of-the-art method can not precisely deal with objects at the edge of pictures due to the lack of contextual information,this thesis proposes a contextual information fusion counting network based on local regression.Inspired by the focal loss designed for classification task,this thesis constructs a focal regression loss to alleviate the problem of sample imbalance during model training.The experimental results on maize tassel count data set verify the effectiveness of the network structure and loss function in this thesis.In addition,this thesis also introduces some state-of-the-art object detection and counting methods,and comprehensively evaluates and analyzes them on maize tassel dataset to provide a reliable reference for relevant researchers.Second,different from existing methods that treat object counting as a regression problem,this thesis redefines it as classification task by discretizing the count value into different numerical intervals.However,classification tasks are usually supervised by the softmax loss,which just treats all negative classes identically when updating the network.Considering the extra ordinal information between different classes of object counting task,this theis proposes a mean-variance loss as a supplement to the cross-entropy loss to make full use of this order information.By conducting experiments on maize tassel counting dataset,this thesis fully validates the proposed method.Finally,this thesis first considers the domain adaptation problem in plant-related counting task.The computer vision model in the agricultural context can only be constructed based on historical data but usually be applied to unknown field scenes.As factors such as variety,year,region lead to the mismatch of data distribution easily,there is a significant performance degradation of the model.To correct this distribution shifts,this thesis proposes an unsupervised domain adaptive local regression model by aligning feature distribution with H-divergence.Taking category as example,this thesis verifies the necessity of domain adaptation and the effectiveness of proposed method by conducting sufficient experiments on maize tassel counting dataset.In summary,taking maize tassel for example,this thesis explores the problem of infield object counting and proposes several computer vision based solution.The work we have done in this thesis is of important theoretical significance and application for counting tasks in agricultural automation. |
