Abstract: According to the invariant moment theory, the binary and normalized maize disease images are obtained. A?new?and?better?RBF-BP?neural?network?recognition?system?with?the?approximation?and?the?fault?tolerance?is?proposed. The Hu invariant moment’s advantages that contain translation, proportion, rotation invariant and good anti-jamming are all used to deal with the complex and changeful maize disease images. According to the invariant moment’s reliability, independence, and little number of those characteristics, it can get a better convergence of recognition system to extract the maize image’s features and the training and recognition of RBF-BP neural network. The results of simulation show that the maize disease recognition of RBF-BP neural network has high accuracy and efficiency.

Key words: maize disease recognition, Hu moment invariant, Radial Basis Function-Back Propagation（RBF-BP） neural network

摘要： 基于不变矩理论，对玉米病害图像进行二值化、图像归一化处理，提出一种新的、具有较好逼近能力和较强容错能力的RBF-BP神经网络识别系统。利用Hu不变矩特征的平移不变性、比例不变性、旋转不变性和对目标良好的抗干扰性等特性，处理复杂、多变的玉米病害图像，形成不变矩特征矢量样本库。根据Hu不变矩在提取图像特征过程中的可靠性、独立性及数目小的特点和RBF-BP神经网络在识别过程中较好收敛性特点，对玉米病害图像进行特征提取、网络训练和病害特征的识别。仿真实验结果表明RBF-BP神经网络系统的有效性。

关键词: 玉米病害识别, Hu不变矩, 径向基函数-反向传播（RBF-BP）神经网络