Plant pests and diseases image recognition is one of the key technologies of digital agricultural information collection and processing. Usually, based on pest infestation-like plant, it is carried out according to the size, shape, color, texture, etc., or a combination of several parameters. Machine recognition of diseases and insect pests needs to use digitalized characteristics without overlapping. Multi-fractal analysis of Fourier transform spectra was adopted to investigate the possibility of extraction of damage pattern characteristics for Citrus reticulata Blanco var. Ponkan. First, images of the boundary of a damaged pattern are extracted with an improved watershed algorithm and region merging. Secondly, a Discrete Fourier Transform (DFT) was applied to the damaged fruit image. With reference to the boundary of a damaged pattern, a fruit image magnitude spectrum was extracted. Thirdly, a fruit image magnitude spectrum was multi-fractiously analyzed and the multi-fractal spectrum of DFT magnitude spectrum was quadratic fitted. Height, width, and centroid coordinate of a fitting parabolic section were chosen feature values to identify the diseases and insect damage of fruits, with these three feature values as inputs of a BP neural network identifying diseases and insect damage of Ponkan, and the accuracy was up to 92.67%. Finally, the amplitude spectrum of the Fourier transform was adopted for multifractal analysis and multi-fractal spectrum of a quadratic fit;fit parabola segment height, width, and centroid coordinates were regarded as pests’ Eigen values, and then used as input variables to establish a BP citrus pest identification neural network model for pest identification. Among 5 classes of pests, in 30 groups of test samples, such as Pezothrips Kellyanus, Oxycetonia Jucunda, Oraesia Emarginata, Polyphagotarsonemus Latus, Colletotrichum Gloeoporioides Penz, the highest recognition rate was for Oraesia Emarginata, that is 96.67%, Polyphagotarsonemus Latus was the lowest at 86.67%, and the average correct recognition rate was 92.67%. The test came to the conclusion that the height, width, and centroid of a multi-fractal spectrum of a Fourier transform spectrum of damaged fruit image better illustrates the features of the disease and insect damage of fruits, such as a complicated biological entity. This method is possibly applicable to automatic recognition of disease and insect damage of Citrus reticulata Blanco var. Ponkan, and it’s able to be applied to disease and insect damage recognition for other plants.%为探讨植物病虫害互不交叉、重叠的数字典型特征值来进行病虫害计算机识别，研究了椪柑病虫害为害状图像傅里叶变换幅度谱的多重分形特征。首先，用改进型分水岭算法检测病虫害为害状边缘，并对其进行区域合并，形成病虫害为害状边界。其次，对病虫害果进行二维离散傅里叶变换，依据病虫害为害状边界进行图像标记，提取标记区域内的傅里叶变换幅度谱图。最后，对傅里叶变换幅度谱图进行多重分形分析及多重分形谱的二次拟合，将拟合抛物线段的高度、宽度和质心坐标作为病虫害特征值，并以此为输入变量，建立 BP 神经网络椪柑病虫害识别模型来进行病虫害识别，椪柑蓟马、花潜金龟子、吸果夜蛾、侧多食跗线螨、椪柑炭疽病5类病虫害30组测试样本中吸果夜蛾识别正确率最高96.67%，侧多食跗线螨识别正确率最低86.67%，平均正确识别率为92.67%。试验结果表明：傅里叶变换幅度谱图的多重分形谱高度、宽度和质心坐标较精确地刻画了病虫害为害状这类复杂生物体的特征，该方法可进行椪柑病虫害自动识别，并可推广到其他植物的病虫害机器识别中。
展开▼
