An Improved Model for Plant Disease Analysis Using Attention Based Capsule Network

The existence of crop disease leads to a substantial loss in quantity and quality in agricultural production.The regular monitoring of plants helps in preventing leaf disease at the early stages and is one of the beneficial acts in preserving the plant.This research focuses on developing an automated approach to classify plant leaf disease using Efficient CapsNet with Self-Attention Routing (EC-SAR).The raw data is obtained from plant village and CGIAR dataset which is pre-processed by color enhancement and data augmentation.Next to pre-processing, segmentation is performed using K-means clustering and Otsu threshold techniques which effectively segment the partitions of leaves.The feature extraction is performed using Gray Level Co-Occurrence Matrix (GLCM), Local Ternary Pattern (LTP) and Histogram Oriented Gradients (HoG) techniques.After this, the classification is performed using the proposed EC-SAR which is comprised of an efficient capsule network that adapts itself to the large input data.The self-attention layer routing is performed to choose a better capsule that is suitable to generate feature representation based on actual data.The experimental outcome shows that the suggested framework achieved an accuracy of 99.98% for the plant village dataset which is comparatively higher than the existing few-shot learning approach, Convolutional Block Attention Module and Efficient Channel Attention (CBAM-ECA) and Loss Fused Convolutional Neural Network (LF-CNN) with accuracies of 93.45%, 99.74% and 98.93% respectively.