Enhancing the Food Image Classification Accuracy Using Ensemble of CNNs and HelperNet

Abstract

Image classification is an important research area spanning diverse fields and has evolved finding new applications with the use of deep learning techniques involving convolutional neural networks (CNNs). With the growing penetration of the social media in day-to-day lives, people sharing food images on various platforms has necessitated improved food classification systems. In this work, we use multiple combinations of the food image classification CNNs and improve the classification accuracy for the food images. We propose a multi-layer CNN architecture called the HelperNet, which improves the accuracy of food image classification when combined with the traditional CNNs like AlexNet, GoogLeNet, ShuffleNet, and SqueezeNet. We show through the simulation results that such a combination gives the accuracy value of 95.5%. We extract the features from the traditional CNNs and combine these multiple features to form a significant feature vector to get a better classification accuracy. We also propose image pre-processing, i.e., sharpening the image center and fading the boundaries, further enhancing the image classification accuracy. Further, we also compare the effect of feature extraction from different levels of CNNs. The analysis and numerical results on classification accuracy serve as a useful benchmark in this burgeoning field.