A Novel Hybrid Deep Learning Framework using Squeeze VGG Vision Transformer Network for Early-stage Disease Detection in Plants

Abstract

According to the historical database there is less quality due to the significant impacts of plant diseases that leads to lack of empowerment for framing as well as growth in agriculture. Early-stage detection is still challenging while identifying symptoms such as minor injuries, discolorations in leaves, major texture variation on leaf surfaces and so on. Convolutional Neural Network has a more effective framework that captures local texture patterns where there is still a lack of capabilities in existing models. A long-range spatial communication between each leaf pattern is the main problem that was addressed using Squeeze-VGG Vision Transformer Network (SVViT-Net) highlighted as a proposed novel framework which can predict plant disease in the early detection stage. Integration of VGGnet Convolutional layer for extraction of local features which is fine-grained is done with SqueezeNet model which reduces features to minimal that can reduce computational complexity. The minimal feature maps which are extracted are then transformed into patch embeddings and then fed to Vision Transformer to identify global contextual dependencies with the help of multi-head automated attention techniques. PlantVillage dataset from kaggle are involved for experimental purposes where 54,000 images of healthy plants from various categories as well as diseased leaves from different specimen crops are used. Initially images are pre-processed by resizing and then normalized as per necessary features with data augmentation steps which can enhance model generalization. Results are implemented using the images and proven with Adam optimizer as evaluation of various metrics using SVVITNet architecture which achieved 98.1% of accuracy, 97.8% of precision, 97.6% of recall and 97.7% of F1-score which outperforms bestfit model as integration of vision transformer and EfficientNetB0 which can be involved for crop disease monitoring.