An Enhanced Plant Disease Classification Model with Novel Architecture for Computational Cost Optimization and Fusing Multi Feature Extraction Techniques

Abstract

This research delves into the intricate realm of plant disease classification, employ- newlineing advanced image processing and machine learning techniques to address this critical newlineagricultural challenge. The study encompasses three contributions, each meticulously newlinedesigned and evaluated to enhance disease detection accuracy. newlineThe initial contribution underscores the significance of feature extraction methods. newlineThe Grey-Level Co-occurrence Matrix (GLCM), statistical features, and wavelet trans- newlineform are harnessed to extract salient attributes from plant images. A novel modification to the GLCM technique is introduced, isolating leaf features from background elements. The Synthetic Minority Over-sampling Technique (SMOTE) addresses dataset imbalances and prevents bias that may affect model performance. Combining these feature extraction techniques and SMOTE improves classification accuracy, effectively addressing classification challenges. The second contribution introduces a novel approach called Binary Feature Map-Splitting Architecture (BFMSA) to improve performance newlineand reduce computational costs. The contribution is two-fold: building a CNN model from scratch based on BFMSA reduces the trainable parameters in both the feature newlineextraction and classification phase. In contrast, applying BFMSA to different transfer learning models reduces the trainable parameters in the classification phase. Moreover,the proposed architecture reduces the number of trainable parameters by up to 87% compared to traditional models. newlineThe third contribution proposes a novel approach for extracting deep features of newlinediseased plant leaves. The proposed approach is based on a Transfer Learning-based newlinemodel followed by a vision transformer (TLMViT). The main contribution is extracting newlinethe deep features of the leaf using two consecutive phases. A pre-trained model extracts initial features in the first phase, and the feature maps of initial features would be fed to ViT for deep feature extraction in the second phase. TLMViT performs accurately

Description

Keywords

Citation

item.page.endorsement

item.page.review

item.page.supplemented

item.page.referenced