ATTNLABNETV3+ AND PLANTSHIELDNET: AN OPTIMIZED DEEP LEARNING APPROACH FOR ACCURATE TOMATO LEAF DISEASE CLASSIFICATION

Abstract

Accurately identifying tomato leaves in the field is essential for estimating yields early on. It is challenging to identify which diseases are specifically affecting tomato plants due to the overlap of similar symptoms of several diseases. Though Deep Learning (DL) algorithms aid in accurate classification, there is room for improvisation concerning computational complexity, and minimization of false positives. For this reason, this research intends to introduce a novel tomato leaf disease classifier with the objective of accuracy maximization. This procedure begins with pre-processing the raw tomato leaf images to make them fit for further processing. It includes image enhancement using CLAHE to improve contrast, data augmentation (random flips, rotations, zoom, and brightness adjustments) to increase dataset diversity, resizing and Z-score normalization to resize images to the required input dimensions of the DL model and normalize pixel values, and Non-Local Means (NLM) denoising for noise removal. In order to find the diseased region, the Region of Interest (RoI) is identified using proposed AttnLabNetV3+. Once the diseased region is segmented, the disease is identified using the proposed PlantShieldNet classifier which integrates DenseNet121, InceptionV3, ResEffi-Net, and Spatio-Temporal Attention. For further improvement, the hyperparameters of the proposed PlantShieldNet are tuned via the proposed Adaptive AdaBelief Hyperparameter Optimization (AAHO). The proposed classifier attained promising results of 99.25% accuracy and outran other SOTA models.