基于无人机可见光图像Haar-like特征的水稻病害白穂识别

实现稻田精准植保的关键是自然环境下病变区域的准确识别。为实现大面积稻田中白穗的精确识别,该文提出一种小型多旋翼无人机水稻病害白穂识别系统,该系统以无人机平台作为图像采集、处理和识别的基础,首先对白穗图像提取Haar-like特征,其次以Adaboost 算法进行白穗训练识别。以4类Haar-like特征及其组合构建弱分类器,用采集的稻田白穗和背景共700个样本点训练生成强分类器。所得强分类器对测试集中65幅图像中的423个白穗样本点进行识别验证,结果表明：白穗识别率可达93.62%,误识别率为5.44%,该方法可有效抑制一般的稻叶遮挡、稻穗黏连以及光照等复杂背景的影响,适合于自然环境下的稻田白穗现场识别。

Identification of diseased empty rice panicles based on Haar-like feature of UAV optical image

Empty rice panicles are a common pest and disease characteristic in rice fields that affects the rice yield and quality. In order to achieve accurate prevention and control of pests and diseases in rice fields, in this study, a multi-rotor UAV-loaded industrial CCD digital camera was used as the image acquisition platform to rapidly and accurately identify and locate the empty rice panicles in large area rice fields based on the Haar-like feature extraction and Adaboost training algorithm. We used the method of UAV aerial photography technology to perform video capture of large area rice fields on a scheduled route. The interval frame number of the sample image was calculated by parameters such as the flight speed of the UAV, aerial video speed, aerial altitude, and the angle of camera, then the video of the rice field was processed by image disassembly, frame extraction, image mosaic, etc. to achieve efficient and rapid acquisition of image information of large area rice fields. The training sample database and the test sample database for the test were finally formed according to the position information of the rice field coordinate in the frame image extracted by the image extraction interval frame number. After many preprocessing operations, such as compression, cutting, normalization, background separation, threshold segmentation, noise removal, etc., the images in the training sample database and the test sample database were applied in the Haar-like feature extraction and AdaBoost training. In this study, we designed four kinds of Haar-like features, such as edge feature of class A, linear feature of class B, center feature of class C and extension feature of class D, these four kinds of Haar-like features and their combination features were rapidly extracted by the integral diagram calculation, then input the extracted Haar-like features into Adaboost training. During the calculation process, we took each discrimination threshold based on the Haar-like features as a weak classifier to give iterative cycle training, after T times iterative cycles. Then T weak classifiers were obtained, and the strong classifier was obtained after cascading the weights of the T weak classifiers. After the Adaboost training, the obtained strong classifier minimized the misjudgment rate of weak classifiers at all levels in each cycle of iteration. We then took the Haar-like eigenvalue extracted by the unrecognized samples as the input of the strong classifier, based on eigenvalue weight, the strong classifier gave a assessed value H to judge whether it was the empty rice panicles or not. When the H was 1, it meant that the classification result was empty rice panicles. When the H was -1, the tested sample was not the empty rice panicles. In this way, identification of empty rice panicles was realized. In order to ensure the diversity and adequacy of the test samples, the influence of the interference factors such as various forms of the empty rice panicles, lighting, shielding, adhesion and background etc. were fully considered. Two hundred and eight five images and a total of 700 positive and negative samples in the training sample database were used for Haar-like feature extraction and AdaBoost learning training. Sixty five images and a total of 800 positive and negative samples in the test sample database were used to verify the performance of strong classifier. The experimental results showed that among the four Haar-like features and their combined features, the class C and class D Haar-like combined features had better performance in improving classifiers than other features. The strong classifiers generated by this combined features were then used to identify the 423 empty rice panicles samples in the test, among which, three hundred and ninety six were identified, and the recognition rate was 93.62%. Our results demonstrated that this method could effectively inhibit the influence of complex backgrounds such as the rice leaves shielding, rice panicles adhesion and lighting etc., and it was also suitable for field identification of empty rice panicles in natural environment. In the study, this method was compared with algorithms that used texture recognition, such as shear waves, contour waves, curve waves, etc. The experiment showed that this method has significant advantages both in the accuracy and the speed of recognition.