Serial Fusion based Hybrid Features Extraction Method for Tomato Plant Leaf Disease Detection

Abstract

Detecting and classifying leaf diseases at an early stage is crucial. However, many current approaches have struggled to achieve the necessary accuracy due to several challenges. These challenges include dealing with input images that have low contrast, irregularities in leaf spots, minimal color differences between foreground and background, and the need to handle a large number of extracted features. To address these issues, we have explored feature extraction techniques based on color, shape, and texture, each with its own advantages and disadvantages. We have conducted experiments on color histograms, color means, color DHV, dominant colors, and obtained results. In shape feature extraction, we have experimented with existing methods such as linear, UNL, Gaussian, grayscale Fourier, area, perimeter, roundness, compactness, and convex hull parameters. The results obtained from color and shape feature extraction methods show that combining multiple features of the image to detect leaf infections provides better accuracy compared to using single feature types. To extract features from the images, we have proposed a hybrid (serial fusion) strategy for color, texture, and shape feature extraction, which achieves the desired accuracy.