基于为害状色相多重分形的椪柑病虫害图像识别

为自动识别椪柑病虫害,研究了以椪柑病虫害为害状多重分形谱特性参数为输入变量的小波神经网络病虫害识别方法。利用改进型分水岭算法提取椪柑病虫害为害状边界,对非连续的边界进行边界跟踪,将过分割区域进行区域合并,标记为害状边界,提取标记区域,生成病虫害为害状目标图像;对病虫害为害状目标图像0°~120°这一主要色相区域4等分,产生4幅色相二值图像;对二值图像进行多重分形分析,计算其标度不变区多重分形谱的高度及宽度;以此高度及宽度作为小波神经网络的输入,进行椪柑病虫害识别,5种病虫害的平均识别正确率为87%。试验结果表明:椪柑病虫害为害状的4对多重分形谱高度及宽度值较充分地反映了椪柑病虫害色相累计信息、分布信息及区间形状的典型特征,能用此方法进行椪柑病虫害机器识别。

Damage Pattern Recognition of Citrus reticulate Blanco Based on Multi-fractal Analysis of Image Hue

The investigation proposed a new algorithm to automatize the identification process of pests and insects disease of Citrus reticulata Blanco var. Ponkan, in which multi-fractal spectra of image hue were set as inputs of wavelet neural network model. In the new algorithm, image boundary of damage pattern of Ponkan was extracted with improved watershed algorithm, and discontinuous boundary was processed with boundary following, meanwhile over-segmentation region was merged and boundary was marked, at last, damage pattern image was generated. After the work above, firstly, hue range 0°~120° of damage pattern image was equally segmented into 4 regions to generate 4 binary images. And then these binary images were analyzed by multi-fractal method to calculate the widths and heights of multi-fractal spectra of scale invariance region. In the end, the widths and heights of multi-fractal spectra were set as the inputs of wavelet neural network model to identify the pest and insects disease of citrus fruit. Test results showed that the accurate rate of identification of 5 pests and insects disease is about 87%, which means that widths and heights of multi-fractal spectra are sufficient to characterize the damage pattern of citrus fruit, and this method is applicable in machine automatic recognition for pests and insects disease of citrus fruit.