Research On Fast Detection And Identification Of Field Pests Based On Machine Vision

The fast detection and recognition of pests is the basis of crop pest and disease control in agriculture. Traditionally, experts observe the external features of pests and then compare these features with specimens while identifying pests, which is time-consuming and labor exhaustive. With the development of computer technology, image processing technology and pattern recognition technology were used to the research and identification of pests gradually and the identification system of pests was established, which not only enriched the means of identification but also improved the efficiency of identification.Typical agricultural pests were studied as research objects in this thesis. Image segmentation, feature extraction, classification, etc. were studied based on pests’ images with the technologies of digital image processing, pattern recognition. A remote automatic insect recognition system based on internet of things was established based on these results and technologies through3G wireless network. The main research contents of thesis are as follows:(1) The image acquisition system was designed. The length of pests in this research have great difference, the same pest always has different attitude and pests have strong activity ability. Therefore, the pest image acquisition systems which meet two kinds of demand were designed. The first system was designed to get image of trapped pests which were stationary. The distance of pests from the system and the scope are fixed. Therefore, the camera uses the mode of CMOS line scan and the focal length of camera lens is fixed. The second system was designed to get pest images in the field real-time. The pests were active, the distance of pests from the system can change and the scope is fixed. Therefore, the camera uses the mode of CCD line scan and the focal length of camera lens is adjusted.(2) The segmentation technology of pest images based on HSV color model was put forward. According to the characteristics of the sample background and objectives, the Otsu threshold segmentation method based on HSV color model was applied. The RGB model was transformed to HSV model before segmentation and the Otsu algorithm was applied to H component of pest images for getting the threshold adaptively. To get the whole pests, the subsequent processing was done after finishing the segmentation. The applied method overcame the disadvantages that many backgrounds were misclassified as target when using the pest RGB image. (3) Multi-feature extraction and feature selection technologies of pest images were studied. The visual features including the morphological characteristics, color characteristics and texture feature were extracted and the redundant features were eliminated through the Ant colony optimization algorithm(ACO). With the ACO, the35dimensional features were lower to29dimensions and the recognition accuracy was improved from87.4%to89.5%. The SIFT algorithm which is considered easy to use by all in recent years was applied to extract local features of pests in this study. The local features are invariant to image scaling, translation, rotation, partially invariant to illumination changes and affine or3D projection and independent on the background segmentation which are suitable for the extraction of pest image features obtained in natural light and complex background. The application of the local feature in the identification of pests not only expanded the application field of the local feature but also provided new ideas and new methods for pest feature extraction.(4) The method of pest pattern recognition was studied. The SVM model was used to identify the pests. The theory of foundations and basic methods of the support vector machine(SVM) were elaborated. The different identification experiments using different pest features were done and the test accuracies were compared in this thesis. The accuracy using morphological characteristics, color features and texture features was over85%. The accuracy using the local features obtained by SIFT algorithm in original RGB images was79%. The test results showed that the ant colony optimization algorithm can eliminate the correlation of image feature and improve the accuracy. At the same time, the test results showed that the local feature extraction method can be applied to obtain the features of pests with no background segmentation.(5) The pest remote automatic identification system based on internet of things was designed. The thesis took twelve kinds of typical agricultural pests including Cnaphalocrocis medinalis Guenee, Prodenia litura, Chilo suppressalis, Sesamia inferens, Anomala corpulenta Motschulsky, Ostrinia nubilalis, Naranga aenescm, Sogatella furcifera, Agrotis ypsilon Rottemberg, Gryllotalpa orientalis Burmeister, Diaphania perspectalia and Conogethes punctiferalis as the object of study. The system included one host control platform and more remote platforms which formed a distributed identification network through3G wireless network. The identification process can be finished in remote automatically and can also be done in the host control platform after the pest images were compressed and transmitted to the host control platform through3G network. The system has the function of expanding the sample library dynamically by getting the image from the local disks. The system also included the expert identification interface. The expert can identify the pests which have been classified by the system automatically and can compare the identification results through the interface. Because of adopting the pest images obtaining in natural light and random pest attitudes, the recognition model has strong generalization ability which is superior to research existing.