Leaf disease detection and classification in food crops with efficient feature dimensionality reduction.

Abstract

Computer vision heavily relies on features, especially in image classification tasks using feature-based architectures. Dimensionality reduction techniques are employed to enhance computational performance by reducing the dimensionality of inner layers. Convolutional Neural Networks (CNNs), originally designed to recognize critical image components, now learn features across multiple layers. Bidirectional LSTM (BiLSTM) networks store data in both forward and backward directions, while traditional Long Short-Term Memory (LSTM) networks handle data in a specific order. This study proposes a computer vision system that integrates BiLSTM with CNN features for image categorization tasks. The system effectively reduces feature dimensionality using learned features, addressing the high dimensionality problem in leaf image data and enabling early, accurate disease identification. Utilizing CNNs for feature extraction and BiLSTM networks for temporal dependency capture, the method incorporates label information as constraints, leading to more discriminative features for disease classification. Tested on datasets of pepper and maize leaf images, the method achieved a 99.37% classification accuracy, outperforming existing dimensionality reduction techniques. This cost-effective approach can be integrated into precision agriculture systems, facilitating automated disease detection and monitoring, thereby enhancing crop yields and promoting sustainable farming practices. The proposed Efficient Labelled Feature Dimensionality Reduction utilizing CNN-BiLSTM (ELFDR-LDC-CNN-BiLSTM) model is compared to current models to show its effectiveness in reducing extracted features for leaf detection and classification tasks.