Transfer learning-based texture-enhanced convolutional neural networks over plant disease identification

Abstract

The global agricultural productivity and food security take serious threats due to the presence of plant diseases; thus, early and accurate diagnosis becomes the key to successful management of the disease. The traditional diagnosis techniques that rely on visual observation are time-based, subjective, and cannot be implemented on a large scale. Recent development in machine learning and computer vision provides possible solutions to automated plant disease detection. This paper suggests a plant disease identification with transfer learning (PDD-TL) model with the preprocessing, segmentation, feature extraction, and disease prediction phases. In the initial stages, median filtering is used to simplify the image quality, after which cells affected by the disease are segmented with the help of the integration of adaptive pixels in joint segmentation (IAPJS) algorithm. Multi-texton and pyramid histogram of oriented gradients (PHOG) are the discriminative features extracted. The classification of the disease is done with a new triple convolutional activation CNN with transfer learning (CNN-TCA-TL). In contrast to the current methods that use either a pure deep learning method or handcrafted features, the framework proposed explicitly employs both the use of texture descriptors and transferable deep representations, which retain fine-grained structural details. The experimental findings prove that CNN-TCA-TL has an accuracy of 0.92 which will prove that it is effective.