基于子图选择的土种识别

针对机器视觉野外自然条件下采集土壤图像包含阴影、空洞、缝隙对土种识别可能产生影响,和土壤图像采集、标记成本高,样本量小的问题,提出一种基于土壤子图选择的土种识别方法。该方法以局部阴影最小化及空间距离最大化为准则,构造最优化模型；并基于最大最小距离算法思想,迭代更新距离矩阵求解优化模型,获得选择子图中心点,构建土壤子图数据集。实验结果显示：土壤子图选择算法构建的数据集在三个不同深度ResNet模型下训练、识别测试,均有较好的训练、识别精度；在土壤子图选择α参数为1、子图尺寸为224时,在浅层网络ResNet-18下能达到最佳测试结果,训练的网络模型在验证集识别准确率为92.48%,测试集识别准确率为92.95%,相较于土壤原图数据集的最佳土种识别结果提升46.65%；利用最小外接矩阵能加速相较于不使用加速,子图选择算法提升运算速度38.88%。基于子图选择的土种识别大幅度提升土种识别准确率,证明算法是有效的。

Soil species classification based on sub-image selection

The presence of shadows, holes, and gaps on soil images collected under natural conditions has a significant negative impact on machine vision classification of soil species. And the sample size for machine vision-based soil species identification is generally insufficient due to the high cost of acquiring and labeling soil photographs. In this paper, we present a soil species classification method based on sub-image selection to address above problems. The proposed approach first constructs an optimal model to select the centers of soil sub-images according to the principles of local shadow minimization and spatial distance maximization. And then creates a soil sub-image dataset on the basis of a center set obtained by iteratively updating a distance matrix utilizing above optimization model. Experimental results demonstrate that soil sub-image datasets established by the sub-image selection algorithm exhibit excellent performance on three ResNet models with different depths. ResNet-18 performs best when set the adaptive factor α to 1 and the sub-image size is 224, with a validation set classification accuracy of 92.48% and a test set classification accuracy of 92.95%. In addition, it is 46.65% superior than the best classification result of the soil species classification model rained with the original soil image dataset. These prove that the accuracy of soil classification based on the sub-image selection algorithm is promoted and the algorithm is effective.