Abstract:

Automatic identification and count of rice light-trapped pests is a common and important pest forecasting method in paddy fields. However, most of the light-trapped pests are unnecessary to be monitored and must be removed. This manual method is time-consuming and fatiguing with a low accuracy rate. We developed an automatic method for identifying rice light-trapped pests based on the images. Firstly, we divided the images into several groups according to their morphological features. Each group has three classifications: the back-side image of a pest, the abdomen-side image of this pest, and non-forecasting pest image similar to this pest. Then, thirty-one features including the color, shape and texture features were extracted from each insect image. Finally, three support vector machine classifiers with posterior probabilities were used to train and test the three groups of insects, respectively. In the results, the back-side image and the abdomen-side image of a pest were considered as the same species. We achieved a 97.4% accuracy rate in the three species of rice light-trapped pests.

Key words: light-trapped pests, image features, support vector machine classifier, automatic identification, non-forecasting pests rejection, forecasting, rice

摘要：

利用灯光诱杀稻田害虫,并识别与计数这些害虫是水稻害虫的一种常规但非常重要的测报方法。在灯光诱杀的昆虫中,大多数昆虫是不需要测报的,因此,在人工识别灯诱测报害虫时,需要排除这些昆虫。这种人工识别与计数害虫的方法效率低、任务重、识别准确率差。我们提出了一种基于图像的水稻灯诱害虫自动识别方法。首先,根据测报害虫的形态特征对水稻灯诱昆虫图像进行初步分组,每组昆虫图像中包含一种测报害虫的背面图像、腹面图像和与其形态相似的非测报害虫图像,共3类;然后,提取组内每一张水稻昆虫图像的颜色、形态和纹理共31个特征参数;最后,利用带后验概率的SVM分类器对每组的3类昆虫图谱进行训练和测试,输出时同一种测报害虫的背面和腹面图像被视为同一种害虫。结果表明,3种较大个体的水稻灯诱测报害虫的平均识别准确率为97.4%。

关键词: 灯诱害虫, 图像特征, SVM分类器, 自动识别, 非测报害虫排除, 测报, 水稻