Deep Learning and Optimization-Based Pest Detection in Peanut Crops Using CNN, MFO, and EViTA

Abstract

The rapid advancement of Vision Transformer (ViT) methods has proven highly effective in image classification and identification tasks. This paper introduces an Enhanced Vision Transformer Architecture (EViTA) tailored specifically for pest identification, segmentation, and classification. Building upon ViT's strengths over Convolutional Neural Networks (CNNs), EViTA aims to improve accuracy in pest image prediction. The methodology incorporates preprocessing techniques such as Moth Flame Optimization (MFO) for image flattening and normalization, along with a dual-layer transformer encoder for integrating pest image segments of varying sizes. Extensive experiments using three pest datasets affecting peanut crops demonstrate the efficacy of EViTA, achieving promising results. Furthermore, the exploration of additional techniques such as DenseNet, InceptionV3, and Xception TL models suggests potential accuracy improvements beyond 94%. Additionally, the integration of Flask framework enables the development of a user-friendly front end for testing with authentication. EViTA presents a novel approach to pest identification with significant implications for enhancing pest management and agricultural practices. Further research and refinement hold promise for advancing EViTA's capabilities in pest identification tasks.