One of the world's most widely grown crops is corn. Crop loss due to diseases has a major economic effect, putting the food supply in jeopardy. In many parts of the world, lack of infrastructure still slows disease diagnosis. In this context, effective detection of corn leaf diseases is necessary to limit any unfavorable impacts on the yield. This research has been carried out on the corn leaf images, having three classes of diseases and one healthy class, collected from web resources by using the densely connected convolutional neural networks (CNNs). In this work, VGG16, a variant of CNN, is investigated to classify the infected and healthy leaves. We conduct four different sets of experiments using pretrained VGG16 as a classifier, feature extractor, and fine-tuner. To improve our results, Bayesian optimization is used to choose optimal values for hyperparameters, and transfer learning is explored to fine-tune and reduce the training time of the proposed models. In comparison with earlier proven methods, transfer learning on VGG16 produced better results by leveraging a test accuracy of more than 97% while requiring less training time.

中文翻译：

---

使用贝叶斯优化对玉米叶片疾病识别的 VGG16 迁移学习进行超参数优化

世界上种植最广泛的农作物之一是玉米。疾病造成的作物损失具有重大的经济影响，使粮食供应处于危险之中。在世界许多地方，缺乏基础设施仍然会减缓疾病诊断。在这种情况下，有必要对玉米叶病进行有效检测，以限制对产量的任何不利影响。这项研究是使用密集连接的卷积神经网络（CNN）从网络资源中收集的玉米叶图像进行的，该图像具有三类疾病和一类健康。在这项工作中，研究了 CNN 的变体 VGG16 以对受感染和健康的叶子进行分类。我们使用预训练的 VGG16 作为分类器、特征提取器和微调器进行了四组不同的实验。为了改善我们的结果，贝叶斯优化用于选择超参数的最佳值，并探索迁移学习以微调和减少所提出模型的训练时间。与早期经过验证的方法相比，VGG16 上的迁移学习通过利用超过 97% 的测试准确率产生了更好的结果，同时需要更少的训练时间。