Automatic Identification Of Lepidoptera Moth Pests Based On 3D Pose Estimation

Moths are one of the largest groups of Lepidopteran insects,and recognition of species is the basis of prevention and control of pests in agriculture and forestry.Traditionally experts observe the external features of insect and then compare these features with specimens while identifying insects,which is time-consuming and labored.With the rapid development of machine vision and pattern recognition,image-based pest classification and identification has become an important research field,which provides convenience for non-professionals to accurately identify pests.Moth insects,as an important branch of Lepidopteran insects,have a variety of posture forms.The robustness of the pest identification algorithm merely based on two-dimension(2D)image information is negatively affected by the high diversity of insects with multi-posture.The information about 3D pose of moth pest is important parameter to reflect the spatial location of target object.By obtaining the effective 3D pose of moths,the complexity of recognition can be simplified,and recognition efficiency can be improved.It has high application value in the prevention and control in agricultural and forestry pests.In this paper,Helicoverpa armigera(Hubner),Agrotis segetum(Denis et Schiffermuller),Mythimna separata(Walker)were selected as objects of study.This paper presents a method for calculating the angle between the wings of the moths based on the principle of machine vision to determine the three-dimensional posture:The marker feature points of the wings of the moths were extracted by the method of corner detection,spatial coordinates of the feature points were obtained,and then the angle between insect forewing was calculated.The experiment results showed that the relative error was between 0.03%and 3.96%,and the minimum root mean square error(RMSE)value was 1.6533°,and showed that there was no significant difference between the calculated results and the manual measurement by the pair t test.On the basis of the acquisition of the three-dimensional posture information,the texture features,color characteristics and morphological characteristics of the three insects were extracted and regarded as the characteristic parameters of classification.In this paper,the extraction of texture features was determined by the gray level co-occurrence matrix algorithm,Energy,Contrast,Correlation,Entropy were chosen as the characteristic parameters;the color histogram was chosen as the color feature of the image,describing the different color proportion of the whole image;The morphological characteristics of the insect were selected to be simple and effective to represent it morphology characteristics,such as area,perimeter,ration area,stretch,complexity and duty ratio.Based on these,the effective acquisition of the two-dimensional characteristic parameters of moth was realizedAccording to the experimental samples in this paper are small samples,and support vector machine has a better learning performance for finite samples,non-linear high-dimensional problem.The support vector machine classifier was used to classify and identify moths.Through the experiment,we compared the classification effect of selecting different classification features:The recognition rate of classification results obtained by using two-dimensional feature classification of three species was 86.7%.The recognition rate of combining with the two-dimensional characteristic and the three-dimensional pose information was 93.3%.This paper proposed a new pest identification approach based on 3D pose estimation,and the calculated results were consistent with the results of manual measurement,which could provide data of 3D pose.This approach could improve the accuracy of moth pests identification,robustness and it has important significance in the future practical application.Because of the increase of the three-dimensional posture information as identification characteristics,the classification recognition rate has improved significantly.The methods and ideas presented in this paper will provide an important basis for the automatic monitoring and identification of the moth pests in the future.