Integrating Capsule Networks and Bidirectional Gated Recurrent Networks for Effective Classification of Rice Plant Diseases

Abstract

Early detection and precise diagnosis of rice plant diseases during production can minimize damage and protecting the environment and improving yield. Therefore, a workable approach has been required that used to automate the process of categorizing and recognizing disease from images of rice plants. For less computation expense and accurate rise disease classification (RDC), this research can introduce a hybrid deep learning model (DL) that can teamed as Stacked Capsule Position-Attention based Bidirectional Gated Recurrent Network (SCPA_BiGRN) model. The proposed RDC model is composed with four stages including image acquisition, pre-processing, feature extraction and classification. In image acquisition, the sample healthy and unhealthy images are gathered from the publicly accessible datasets including Rice Disease Image (RDI) dataset and Rice Leaf Disease (RLD) dataset. Pre-processing uses Extended Guided Filtering and Removal of Uneven Illumination to lower potential noise and improve the quality of raw input images. For feature extraction stage, Residual-152 model is used to extract the required features from the pre-processed images. Finally, the proposed SCPA_BiGRN model is utilized for classifying the available diseases from the input rice images. In experimental analysis, the proposed RDC model effectiveness is evaluated based on numerous analyzing process namely, performance metrics, quality metrics, ablation study, confusion matrix and learning curve analysis. As the result, the experimental outcomes are demonstrated that the proposed model highly suitable for accurate RDC that can attains 97.42% and 96.33% accuracy in RLD and RDI dataset respectively.