Design of an IoT Based Ensemble Machine Learning Model for Fruit Classification and Quality Detection

Classification of fruit imagery for identification and quality detection of fruits is a multidomain task, which involves pre-processing, segmentation, feature extraction, feature selection, identification and post-processing. Existing models for fruit classification and quality identification use various machine learning and deep learning methods for general purpose processing. Due to which, their accuracy and precision performance is limited. In order to improve this performance for application-specific fruit imagery, this text proposes a novel ensemble learning-based framework, that utilizes ensemble of adaptive Naïve Bayes (aNB), multiple support vector machines (MSVM), varying depth random forests (VDRF), and interpolated k Nearest Neighbor (ikNN) methods. These methods utilize color, texture, shape and convolution features in order to represent input fruit imagery into multiple domains. These feature vectors are given to a variance maximization layer, which assists in feature reduction for reducing classification delay, and improving accuracy of quality prediction. The ensemble model utilizes adaptive feature selection, which is controlled via Genetic Algorithm (GA) based optimization, and assists in identification of optimum model parameters for accuracy maximization. The GA based ensemble model is designed in dual mode, and applied for different fruit image training and validation sets. Due to this, the model is capable of achieving an accuracy of 97.9% for fruit classification, and 93.8% for quality detection. The model is also able to reduce the delay of classification and quality detection when compared with state-of-the-art methods. This is possible due to use of variance maximization for feature selection and classification. The model was further evaluated on multiple datasets, and was observed to be tunable for different fruit types via complete retraining and validation. This characteristic assists in improving scalability of the proposed model, thereby making it applicable for a wide variety of fruit types. This text evaluates the model on Apple and Orange fruits, and recommends steps which must be followed in order to tune the model for any type of fruit for optimum accuracy performance.