为提高基于图像处理的小麦品种识别的准确率,首先选取L8998、内乡188、9023、优展1号、豫麦47、周麦12等6个品种作为研究对象,对采集到的小麦颗粒图像进行中值滤波后采用迭代式阈值法分割图像,提取出颜色、形态和纹理3方面共16个特征,然后通过构建神经网络研究了小麦品种的识别准确率与品种数量之间的关系.最后,为避免网络因达到局部最优而停止训练,利用MIV算法计算了各输入特征参数对分类结果的平均影响值,进而使用遗传算法对网络结构进行了优化.结果表明,随着小麦品种的增加,分类的准确率逐步下降,当待识别的小麦种类增加到6类时,优化后的神经网络的样本识别准确率从81.3％增加至85.6％,有效提高了小麦品种分类的准确性.%To improve wheat variety classification accuracy based on image processing, we selected 6 varieties of wheat samples of L8998, Neixiang 188, 9023, Youzhan 1, Yumai 47 and Zhoumai 12 as objects, then acquired wheat grain images, processed the images by median filtering, segmented the images by iterative threshold method, extracted 16 features from color, morphology and texture aspects, and constructed an artificial neutral network to study the relationship between the classification accuracy and the number of varieties . Finally, we utilized MIV algorithm to calculate the mean impact value of each input characteristic parameter to the classification result to prevent the ANN from trapping into local optimum and stopping training, and further used genetic algorithm to optimize the network structure. The results showed that the classification accuracy dropped gradually when the number of wheat varieties increased; the identification accuracy of the optimized ANN increased form 81.3% to 85.6% when the number of wheat varieties increased to 6, thereby effectively improving the wheat variety classification accuracy.
展开▼
