An Automated System for Disease Segmentation & Classification model via Multimodal Analysis from Thermal Images using Deep Learning Approach

Abstract

Thermal scans assist in representing fruit images as multispectral scans, that can be processed in order to identify different disease types. Existing thermal scan processing models are either highly complex, or cannot be scaled for multiple disease types. Moreover, the efficiency of these models reduces w.r.t. number of diseases, which limits their applicability for real-time scenarios. To overcome these limitations, this text proposes design of a fruit disease segmentation & classification model via multimodal analysis from thermal scans. The proposed model initially collects thermal fruit images, and segments them via entropy-based Saliency Maps. Segmented images are represented into multidomain features via convolution, frequency, entropy, Gabor and Wavelet analysis. These features are capable of representing the segmented images into class-dependent feature sets. To reduce the redundancy from these feature sets, an efficient Whale Optimization Algorithm (WOA) is used, which assists in identification of highly variant features. These features are classified using a binary cascaded CNN (bcCNN), which uses a cascade of multiple binary CNN classifiers. Due to this cascade arrangement, the model is able to identify different fruit disease classes, with high accuracy levels. The cascade binary arrangement keeps the ‘normal’ category common across different evaluations, and uses mode operations to identify output disease class. The model was tested on Apple & Chiku images, and was observed to have 8.3% higher accuracy, 3.9% higher precision, 2.5% higher recall, and 4.9% lower delay when compared with existing thermal image classification methods