Comparison of CNN Tolerances to Intra Class Variety in Food Recognition

Intra-class variation defines image variations occur between different images of one class. The similarity between samples within the same class is typically measured by the Intra-class Correlation coefficient. A high Intra-class Correlation Coefficient close to 1 indicates high similarity between samples from the same class where a low ICC close to zero means opposite. This paper deals with intra-class variety problem of Kegels Foodl0l dataset. 21 classes that have high ICC values were chosen. We have applied well known convolutional neural networks including ResNet, GoogleNet, MobileNet and VGG-Net with different train and test percentages in order to compare the recognition rates for the classes. Although the samples in Food101 dataset vary widely, GoogleNet (Inception V3) has the highest validation accuracy value with the lowest number of epochs.