Rice Pest Image Recognition Research Based On Support Vector Machine

With the development of science and technology and agricultural technology, prevention and identification of rice pests also been growing concern.In this article is mainly using digital image processing techniques and pattern recognition techniques for rice pest species studied classification, and the main contents and results are as follows:(1)Rice pests image preprocessing techniques, we propose a new method for the gray top hat and traditional techniques of integration of the gray-scale image processing, after a lot of experiments, the paper presented a traditional paper new methods in the integration of technology gray top hat processing technology can solve the image during shooting due to the interference problem caused by light and shade, and the traditional gray-scale approach, compared with a good anti-noise performance and robustness.(2)In terms of image de-noising, we propose a filter to automatically determine the radius of de-noising algorithm automatically load an image and recognition by filtering radius of each image for image denoising, filtering radius can be automatically determined by adaptive for each image denoising optimal performance, through a large number of experiments, the proposed algorithm automatically determines the de-noising filter radius and traditional de-noising algorithm, with a matching high, good de-noising performance, so that each one can achieve optimal image denoising radius, and avoid artificial filtering matrix size is set, the program reduces the time complexity and space complexity.(3) In terms of image segmentation, this paper presents an approach based on a single-pixel RGB color background segmentation approach for image segmentation process by identifying a point pixel RGB components of any one component of the time is not 0, then it will become a little black dot, is used to represent an image which occupies storage space on a pixel, so that the amount of calculation can be smaller, without losing large image characteristic parameter information. Compared with traditional segmentation methods, the proposed method avoids image segmentation detail is lost in the process, and other issues of image feature fracture and a large number of holes in the image.(4) Feature extraction pests in rice, taken HSV color feature extraction methods based on histogram ways to improve the identification intuitive; extract morphological characteristics, to ensure that the rotational invariance, image translation invariance and scale invariance. making recognition feature parameters extracted more stable; the use of run-length extracted texture features, you can reduce the workload and reduce the complexity of the calculation procedure.(5) In the feature data dimension reduction using PCA approach to data reduction, this paper presents the difference between the size of a single vector and vector data between the average for dimension reduction, and traditional PCA vector contribution value drop dimensionally improved compared drop dimensional coefficient, the conventional dimensionality reduction contribution value set down dimensional coefficient was 85%, while the paper, based on the vector difference between the way through dimensionality reduction experiments, drop-dimensional coefficient can reach 90% to 95%; compared with traditional PCA approach, this paper presents an improved PCA approach can significantly reduce useless feature, eliminating the correlation characteristics.(6) In rice pest identification, this paper proposes a one-elimination recognition mode, which can identify the different types of multiple samples at the same time, breaking the traditional single-species recognition mode; uses to create multiple recognition classifier image one by one to determine the identification process of elimination, through a large number of experiments show that one of the proposed elimination of identification and compared to traditional means of identification, can improve the recognition accuracy of about 10%, is difficult to avoid the traditional way selection of the optimal kernel function and algorithm stability, high accuracy.