Study On Content-Based Image Retrieval Technology Of Crop Pest And System Development

Text is often applied to describe the information of crop pests and diseases, such as color, shape, etc. in traditional crop disease and pest database or diagnosis expert system. However, text description is inaccurate and subjective, which incurs deviation or even error. Aiming at these problems, this dissertation takes crop pests as research objects, and studied content-based image retrieval methods based on combination of computer vision and database technology, among which image feature extaction, image similarity measurement and user relevant feedback technology are stressed. In the end, content-based crop pest image retrieval system is developed in order to provide technology supports for diagnosing and recognizing crop pests quickly, and sharement of crop pest images. The main contributions of this research include:(1) A new image blocking method is prompted—image is segmented into four overlaped blocks. According to the new segment method, new blocked color moments are proposed to solve the problems that color histogram does not contain the information of spatial distribution of colors across an image. Color moments are used to describe color features together with color layout decribor. As pest images are complex and their boundaries are incomplete, area Zernike moments are put forward to extract image shape features. Features uniform methods are studied and linear weighting among different features is employed to multi-feature image retrieval, by which different features can complement each other to improve retrieval performace. Experimental results present that multi features can better describe image content compared with single features, and their average precision rate is 40.7% and 58.5% respectively, average recall rate is 51.7% and 76.5%.(2) Due to the differences between people's similarity apperception of images and actual distance measurement method, two new methods are prompted to measure similatiy of images—BP neural network improved by parallel simulated annealing genetic algorithm and grey relevant analysis. Experimental results prove that these new approaches return satisfactory results. The average precision rates of distance measurement, improved BP neural...