Comparison of Deep Convolutional Neural Network Architectures for Fruit Categorization

Abstract

In general, there are so many types of fruit images that it is difficult for humans to differentiate them based on their visual characteristics alone. This research focuses on identifying and recognizing images of fruit from 23 different classes or types. Fruit varieties consist of 13 apple classes, 1 orange class, and 9 tomato classes, totaling 15,987 images. Fruit image data were collected from various sources, including the internet, magazines, and direct capture with a digital camera. The process of identifying and recognizing fruit images involves the classification of fruit images using a deep learning algorithm. Several CNN models, which are derivatives of deep learning, are used to achieve high accuracy and robustness in recognizing various types of apples and tomatoes. To evaluate the performance of each model, the apple data were trained on a large and diverse set of apple images using several CNN models such as ResNet50V2, InceptionV3, InceptionResNetV2, VGG16, VGG19, MobileNetV2, and EfficientNet. Performance is assessed using metrics such as accuracy, precision, recall, and F1 score. To achieve optimal performance in the image recognition process, it consists of preprocessing strategies, data augmentation, feature extraction, and classification supported by optimization, all of which have a significant impact on increasing accuracy performance. Experimental results show that certain CNN model architectures outperform other model architectures in terms of time efficiency and accuracy in recognizing fruit types/classes. However, to get more optimal results regarding the performance of the CNN model architecture for fruit categorization, two optimizers will be used, namely Adam and Adagrad, and will be compared. Based on Adam's optimizer experiments, the EfficientNet model produces the highest average accuracy of up to 99%, followed using the VGG 16 and ResNet V2 50 models, which achieve 98% and 97% accuracy. Meanwhile, the use of the Adagrad optimizer with the VGG 16 model produces the highest average accuracy of up to 95%, followed using the VGG 19 and EfficientNet models, which achieve accuracy of up to 93% and 91%. Overall, this experiment produced very good accuracy because it produced an average of above 90%. However, there is still room for improvement in recognizing fruits of different shapes, textures, and colors.