基于4种不变矩和BP神经网络的稻飞虱分类

针对稻飞虱远程实时识别采集图像质量不高的问题，研究了基于不变矩提取形状特征值对稻飞虱进行分类。采用自行设计的拍摄装置采集稻飞虱图像，进行灰度化后用大津法二值化，再用数学形态学滤波；对二值图像采用Hu矩、改进Hu矩、Zernike矩和Krawtchouk矩4种不变矩分别提取特征值，再用BP神经网络进行训练和测试，以此检测4种矩的提取效果。试验用Matlab2008验证算法，对白背飞虱、褐飞虱和灰飞虱共300个样本进行了训练和测试，结果表明Krawtchouk矩提取稻飞虱图像形状特征值的识别率最高，总体达到了91.7%。该文可为大田中现场识别稻飞虱提供参考。

Classification of rice planthopper based on invariant moments and BP neural network

Abstract: Aimed at the problem of the quality of images that were acquired by rice planthoppers remote real-time recognition system, the shape feature values which were extracted by invariant moments to recognize a rice planthopper. 160W self-ballasted high-voltage mercury lamp was used in the experiment to lure rice planthoppers to the curtain, then a H-shape mobile photographing device which had been designed independently by us was used to photograph the planthopper image. The device has the advantages of simple structure and low cost. The USB interface camera of this device was less than 600 RMB. It will lay the foundation for the development of a rice planthopper scene recognition system with low cost. The color images which had been photographed were grayed with a weighted formula, and then were subject to binaryzation with an Otsu method. Finally, the algorithms such as morphological operations were used for filtration to get a binary image with better quality. The feature values of the rice planthopper binary images were respectively extracted by four invariant moments: Hu moment, improved Hu moment, Zernike moment, and Krawtchouk moment, and then a BP nerve network was used to train and test the four feature values respectively, so as to detect the recognition effect of extraction feature values of the four moments. Matlab 2008a was used in the experiment. 240 samples of sogatella furcifera, nilaparvata lugens, and small brown planthoppers had been trained, and then an additional 60 samples were selected for testing. The test result was that the overall recognition rate of the Hu moment was only 76.7%, and the recognition rate of the improved Hu moment was 85%, while the recognition rate of the Zernike moment was 86.7% and the recognition rate of the Krawtchouk moment was 91.7%. The recognition rate of the Krawtchouk moment was the best of the four moments. The reason was that the Krawtchouk moment not only reflected the global feature, but exhibits better locality. The experimental result showed that the Krawtchouk moment has the highest recognition rate. It can be used for the extraction of rice planthopper feature values in the real-time system. This study focused on the search of invariant moments to extract good feature values, but the use of a BP neural network classification resulted in a recognition rate of sogatella furcifera and nilaparvata lugens that was not very high. The identification of sogatella furcifera and nilaparvata lugens was worse than that of the small brown planthoppers. It meant that recognition of two kinds of planthoppers based on a BP neural network needs further study.