An Evolutionary Optimization of Positional-Aware Dual-Attention and Topology-Fusion Generative Adversarial Network for Plant Leaf Disease detection

Abstract

The classification of crop leaf diseases is the foremost essential task in agricultural activities since it may affect the crop productivity. To achieve this task, a Positional-aware Dual-Attention and Topology-Fusion with Generative Adversarial Network (PDATFGAN) can create super-resolution images of crop leaves robustly. Also, Deep Convolutional Neural Network (DCNN) can classify these enhanced images into different types of leaf diseases. But, the adversarial learning objectives can have non-convergent boundary sets near equilibrium which reduces the generative efficiency. Therefore this proposes a new model called PDATF-Evolutionary GAN (PDATFEGAN) by using different objectives to equally optimize the generator and create the super-resolution images for classification. In this model, an EGAN is constructed which considers an adversarial learning process as an evolutionary problem. A discriminator can act as the atmosphere and a population of generators evolve related to the atmosphere. During every adversarial iteration, the discriminator is learned to identify actual and bogus image samples. Also, the generators who act as parents execute different mutations to produce the offspring and adapt to the atmosphere. To decrease different losses between the created distribution and the image distribution providing to various mutations, different adversarial objective functions are considered. Then, the quality and diversity of images generated by the updated offspring are computed for an optimal discriminator. After that, a weakly-conducting offspring are rejected and the residual well-conducting offspring are preserved according to the idea of survival of the fittest for further learning. Further, the generated super-resolution images are fed to the DCNN to identify and classify the types of leaf diseases. At last, the test analysis shows that the PDATFEGAN achieves better accuracy than the existing models.