Research And Development Of Content-based Image Retrieval System With Weed Seed

Weeds have been a threat to the security of agriculture and forestry. To prohibit and restrict the weeds from the source, identifying and quarantining weed seeds quickly and accurately is important for plant quarantine, weed control and agricultural production. The process of traditional identification of weed seeds is slow and inefficient. So in this paper, the content-based image retrieval technology achieves the identification of weed seeds. But most of the existing content-based image retrieval systems is universal retrieval system, which is not optimized for the characteristics of weed seed image. And the retrieval system for a special field is only for legume weed seed images, which could not meet the needs of all weed seeds image retrieval.In view of the problems of above content-based image retrieval systems, features are extracted with the characteristics of weed seed images in this paper. Content-based image retrieval system with weed seeds is designed and implemented to identify weed seeds automatically. The main work in this paper is reflected in the following aspects:(1)Preprocessing of weed seed images. In order to avoid the interference of some noises in weed seeds image, original weed seed images need to be multi-angle normalization and multi-scale normalization before extracting features. Calculating the rotation angle of weed seed images uses the PCA algorithm. Then rotate the image according to the rotation angle, so that the direction of the long axis of weed seeds perpendicular to the horizontal direction, and unify the size of the rotated images.(2)Feature extraction of weed seed images. Some experiments with SIFT, SIFT, ORB, FREAK, BRISK, SURF and HOG are used to extract features for weed seed images. This paper selects SIFT as the feature extraction algorithm, and uses BOW model to optimize SIFT features. The experimental results show that best retrieval results occur when the size of visual dictionary is 200 and the distance measurement of calculating SIFT-BOW histogram is euclidean distance. The average rank of retrieval(AR) is 0.5405, and the average noemalized modeified retrieval rank(ANMRR) is 0.5547. That means that right images account for 54.05% in the front K retrieval results, and 55.47% right images achieve high rankings.(3)Design and implement of content-based image retrieval system with weed seeds. The system is mainly implemented for retrieval and browsing seed weed images. It uses C/S structure, which consists of the client subsystem by the server subsystem. The web service is developed by WCF. The client calls the WCF service layer interface to communicate with the web service. The design pattern of the system is WPF/MVVM which is divided into three layers of Model, View, and View mode. This pattern makes full use of WPF data binding. Three layer structure separates the user interface logic, presentation logic and state control, data and business logic. And the user interface of WPF is designed by XAML language, which can simplify the process of design UI.