LSCEC: Design of a High-efficiency Leaf-image Segmentation & Classification Model via Ensemble Compute Processes

Abstract

Real-time leaf image segmentation requires consideration of shadow effects, light variations, capture-angle variations, and zoom effects. Designing an efficient segmentation model that is capable of considering these effects is highly complex, and requires context-specific segmentation techniques. Such techniques are computationally complex, and require larger delays, which limits their applicability for real-time use cases. Moreover, such models are highly context-sensitive and require continuous reconfiguration for multiple image types. To overcome these limitations, this text proposes design of a high-efficiency leaf-image segmentation & classification model via ensemble compute processes. The model initially uses saliency maps for background removal, and cascades it with an ensemble of multiple Fuzzy C Means (FCM) segmentation techniques. These include, Enhanced FCM, k FCM, and original FCM, which results in 3 different image sets. Each of these image sets are individually processed to identify green & yellow coloured foreground pixels. The extracted image mask is filled for better segmentation, and resulting image sets are processed for extraction of features. To efficiently represent the images, a set of Colour, Texture & Convolutional feature maps are extracted, which assists in identification of multidomain feature sets. These feature sets are classified by a combination of Naïve Bayes (NB) Multilayer Perceptron (MLP), Logistic Regression (LR), and Support Vector Machine (SVM) based classification engines, which assists in improving efficiency of disease detection for different crop types. Due to this combination, the proposed model is able to achieve 3.5% better classification accuracy, 1.8% faster response, 4.9% higher precision, and 5.5% better recall when compared with state-of-the-art models. Due to which, the proposed model is capable of deployment for a wide variety of real-time use cases.