Deep learning technique for grape leaf disease classification via deep high attention stage-by-stage forward Taylor network

Grape leaf disease is a condition caused by various pathogens that lead to discoloration or deformities. Existing techniques detected diseases only after the appearance of visible symptoms, which led to late treatment and reduced crop yields. Therefore, a Deep High Attention Stage-by-Stage Forward Taylor Network (HASFT-Net) approach is developed to classify grape leaf disease. Images of grape leaves were sourced from a database and subjected to a denoising process employing the Attention-Guided Convolutional Neural Network (ADNet) technique. The affected regions are segmented by Spatial Pyramid-oriented Encoder-Decoder Cascade Convolutional Neural Network (SPEDCCNN) with Tversky loss function, and the segmented outcome is trained by the Siberian Tiger Optimization (STO) to enhance the segmentation results. The Opponent Color Local Binary Pattern (OCLBP), Scale Invariant Feature Transform (SIFT), and Shape features (rectangularity, eccentricity, and circularity) are extracted. The grape leaf diseases are classified by HASFT-Net, which strengthens the classification process by integrating the merits of Neuron Attention Stage-by-Stage Network (NASNet), Taylor’s series, and Deep High-order Attention Neural Network (DHA-Net). Results achieved by HASFT-Net are an accuracy of 92.172%, a True Positive Rate (TPR) of 92.883%, a True Negative Rate (TNR) of 91.993%, precision of 90.887%, and F1-score of 91.874% for learning data of 90%.