A joint framework of feature reduction and robust feature selection for cucumber leaf diseases recognition

Abstract

In machine learning (ML) domain, extracted features play a primary role in both segmentation and classification of salient/infected regions. Plants' diseases and pests are the main sources of colossal damage worldwide as they affect both the quality and quantity of crops. In China, cucumber is one of the main crops which is widely cultivated and has high economic benefits. Diseases like angular leaf spot, downy mildew, powdery mildew, etc., affect the cucumber crop and the only method being followed for its prevention is manual inspection. In this work, a hybrid framework based on feature fusion and selection techniques is proposed which classifies the cucumber diseases by using three core steps. Initialized with data augmentation, the contrast of image samples is enhanced in the first step, followed by feature extraction, fusion and selection in the second step. Finally, most discriminant features are classified using a set of classifiers. The extracted features in this work are initially reduced using the proposed probability distribution-based Entropy (PDbE) approach. After the serial-based fusion step, strong features are selected using the proposed Manhattan distance controlled entropy (MDcE) technique which has a capacity to select the features greater than a threshold. From the achieved accuracy (93.50 %) on the selected dataset having more than 900 image samples and six classes, it is very much evident that the presented method is comparable to several other existing techniques.