Clustering of plant disease from digital images is an arduous task due to its dynamic nature and change of appearance under different environmental conditions. In most cases, the image captured in the real-time scenario is subjected to added noise, distortion, poor lighting conditions, and other potential factors that results in poor model performance during the process of discriminating between normal and disease-affected samples. It eventually maximizes the margin of the error rate, thereby leading to misclassification of disease of different varieties of plants in the database with other categories. This paper presents an effective deep clustering-based plant disease categorization algorithm, Boosted-Deep Embedded Regularized Clustering (DEPICT). This model integrates the convolutional autoencoder model with locality-preserving constraints and group sparsity into the network, which improves the embedded learning representation of the images. The PlantVillage and PDD image databases are accessed to develop this model for maize crop. The images are segmented by eliminating the background, cropped, augmented before model training. The performance of the system is evaluated by clustering accuracy and normalized mutual information. The proposed Boosted-DEPICT exhibits better performance, attains promising results with an accuracy of 97.73% and 91.25% on PV and PDD datasets, and outperforms state-of-the-art deep clustering algorithms. This system could be further enhanced by automating the entire process and transforming it into a mobile application for real-time analysis to gain instant results from any region.

由于数字病害的动态性质和在不同环境条件下的外观变化，将疾病从数字图像中聚类是一项艰巨的任务。在大多数情况下，实时场景中捕获的图像会受到额外的噪声，失真，不良照明条件和其他潜在因素的影响，从而在区分正常样本和受疾病影响的样本的过程中导致模型性能不佳。它最终使错误率的裕度最大化，从而导致数据库中具有其他类别的植物的不同品种的疾病分类错误。本文提出了一种有效的基于深度聚类的植物病害分类算法，Boosted-Deep Embedded Regularized Clustering（DEPICT）。该模型将具有局部性约束和群稀疏性的卷积自动编码器模型集成到网络中，从而改善了图像的嵌入式学习表示。可以访问PlantVillage和PDD图像数据库来为玉米作物开发此模型。通过在模型训练之前消除背景，裁剪，增强来对图像进行分割。系统的性能通过聚类精度和标准化的互信息来评估。提出的Boosted-DEPICT表现出更好的性能，在PV和PDD数据集上具有97.73％和91.25％的准确度，并取得了令人鼓舞的结果，并且优于最新的深度聚类算法。