Image Recognition Algorithm For Fruit Flies Based On BP Neural Network

Machine vision is a technology related to artificial intelligence,automatic control,neurobiology and graphics fields of integrated technology.With the development of computer hardware and software and image processing technology,machine vision technology has made great breakthroughs in theory and practice,and has extended to the field of agricultural engineering.At present,the developed countries have used the machine vision technology to realize the recognition and management of the operating objects in the automation of agricultural production.The application research of machine vision technology combined with agricultural engineering in China started late.At the same time,the agricultural production environment in China is complex and variable and unstructured.So the application of machine vision in agricultural production is not mature enough.on the basis of foreign technology,domestic scholars have carried out a lot of research,and applied it to many fields,such as insect quarantine,fruit and vegetable picking,farmland irrigation and so on.It is foreseeable that integrating the cutting-edge technology of computer science into the field of agricultural engineering is a new direction to promote the development of agricultural machinery automation and intelligence in China.In this paper,the dominant species of fruit flies in southern China were taken as the research objects,including three fruit flies of The Bactrocera dorsalis,The Bactrocera tau Walker and The Bactrocera cucurbitae.The images of three kinds of fruit fly adults were processed by computer vision technology and recognition algorithm.The purpose of this study is to identify Diptera fruit fly pests quickly and accurately from images,and to realize process automation by constructing image recognition models,and to provide help for pest control of fruit flies quarantine and wide range crop group real-time monitoring.Fast and accurate identification of pests from images is an important prerequisite for building a real-time monitoring system for insect pests based on machine vision technology.Most of the existing studies use the fruit fly wing structure as a classification feature,which has high requirements for the quality of the fruit flies.If the picture is not clear enough,the structure and wing nevus of the fly wing can not be accurately captured,the recognition effect will be greatly affected.The source images in actual application scenarios may be influenced by external factors such as light and shooting devices,resulting in low quality and affecting the final classification effect.In addition,the most important part of image recognition technology is feature area segmentation.The existing algorithm does not solve the problem of automatically locking feature regions from source images.The source images in actual application scenarios may be limited by shooting angle and fruit fly posture.The accuracy of feature area segmentation can not meet the requirements of recognition algorithm,and the classification results can not meet the requirements.In order to realize a fly classification algorithm suitable for automatic machine recognition,the interference factors must be solved.In order to solve the above two technical difficulties,this paper designs a process for extracting and digitizing the Diptera fruit flies,and encapsulates the process into algorithm module.Combined with the advantage of BP neural network in recognition ability,we build the recognition model.The model can identify the target images in batch,while the efficiency and accuracy meet the real-time requirements.The research work of this paper can be divided into two parts:the characteristic analysis of fruit fly and the construction of image recognition model.(1)Through literature analysis and sample observation,it is found that because of the complexity of the classification of fruit flies,the whole geometric features and color characteristics can not be classified accurately and accurately,and the local characteristics after digitalization processing have better classification characteristics.Based on the mathematical morphology on the common fruit fly,the feature parameters are extracted from the characteristic region with the fruit fly chest backboard as the local area,including the central fringe length R_x,width R_y,eccentricity E,perimeter L,area S,circularity shape parameter C,a total of 7 math morphological features.The statistical analysis of the experimental data was carried out,including the general distribution pattern hypothesis test,the variance homogeneity test and the difference saliency test.The test results showed that the 7 characteristics were generally distributed in the normal distribution of the fruit flies of the same kind.The 4 characteristic values of central stripe length R_x,wide R_y,perimeter L and area S have differences among the three types of fruit flies.The 3 characteristic values of central stripe eccentricity E,shape parameter F,and roundness C are significant differences between The Bactrocera tau Walker and The Bactrocera cucurbitae.Similar stability characteristics and inter class differences make 7 mathematical morphological features serve as the classification basis of fruit fly classification research,and have high reliability,which provides a theoretical basis for subsequent research on fruit fly classification models.(2)In this paper,a recognition algorithm based on neural network learning model is proposed.Using Hough transform and HSV color space,the algorithm is used to automatically segment the feature region of fruit flies.The central fringe of the feature area is further processed.According to the description method of the central stripes,the 4 feature factors of the fruit fly morphology are defined,and the feature vectors are extracted.The process of automatically extracting feature vectors is encapsulated successfully,and the feature vectors of samples are collected in batch as training data set.BP neural network is trained by data sets,and the model parameters for fruit fly classification are obtained.Experimental results show that the recognition effect of this method on Diptera fruit fly adults has good accuracy and real-time.The accurate rate of identification for the Bactrocera dorsalis is 87.3%.The accurate rate of identification for the Bactrocera tau Walker is 86.5%.The accurate rate of identification for the Bactrocera cucurbitae is 87.5%.The total accuracy is 87.1%.Single time identification takes 0.8 s.The further research direction is to increase the dimension of the feature vector and improve the robustness of the algorithm.