A CNN Model-based Ensemble Approach for Fruit Identification Using Seed

With the advancement of computer vision techniques, image classification is being applied in many areas now. Convolutional Neural Network (CNN) has been very successful in image classification and recognition. The traditional method of fruit identification techniques is based on the fruit image itself. This study proposed a CNN model-based ensemble technique for fruit identification that identifies a fruit from its seed image. This study was conducted on 20 different fruit seeds. A basic CNN model was developed primarily. The effect of the number of convolutional layers was examined for the basic CNN model. Transfer learning was performed separately on our basic CNN model by using four pre-trained convolutional neural networks: VGG-16, InceptionV3, Xception, and ResNet as a feature extractor. This helped us to generate four new models. The performance of each model was analyzed. Then we used an ensemble approach by combining those models. Two ensemble models were created. The first ensemble model; which is the fusion of the five individual models, has obtained 88% accuracy while the second ensemble model, which includes four individual models, has achieved 87% accuracy in our dataset. Both ensemble models achieved significantly better scores in other classification metrics: Precision, Recall, and F1-Score compared to the individual models.