Optimized Deep Spectral Generative Adversarial Neural Network (DSGAN2) Framework for Efficient Real-Time Detection of Automatic Rice Leaf Disease

Abstract

Automatic methods which apply computer vision and machine learning to identify rice diseases exists in leaf images, may offer some considerable benefits rather than the manual visual examination methods. In order to detect the rice plant diseases, deep learning techniques are employed, as it supports in lessening the burden of the farmers in securing the crops. Current models are not appropriate for real-time large-scale applications which offer low-resource rural situations; because these current models are often high computational. In this paper we introduced an optimized Deep Spectral Generative Adversarial Neural Network (DSGAN²) to help in improvement of accuracy and reduction in computational complexity. We performed image pre-processing along with gaussian filter or median filter for noise reduction. Moreover, bilinear interpolation is used to resize the image, efficiently. Most relevant features are effectively preserved using feature selection along with the recursive feature elimination. A simplified U-Net model is applied to facilitate an accurate localization; which goes over the excessive resource demand. Feature selection can be additionally optimized by the Binary Particle Swarm Optimization (BPSO). Effective convergence on optimal weights is attained by the BPSO. As we can see from the obtained classification results, False Positive (FP) and False Negative (FN) are minimized, and high accuracy is ensured by an optimised DSGAN², in low-resource conditions. Considerable reductions in error rate at 40.8%, APS-DCCNN at 55.2%, Alex Net at 50.4%, and standard Convolutional Neural Networks (CNN) at 49.5% are obtained by the DSGAN², compared to the current methods. The proposed method brings out an effective robust method for real-time disease detection in agriculture.